RiskMandate v1.41.24 Try it free
Web Summit 2026 · the newsroom map

Everything the newsroom wrote that the site could use.

newsroom.sgit.ai is where the lead has argued RiskMandate in public since August. This is every item that touches it, one row each: its idea in a line, how close it is to what the site sells, what it could become here, and what the article itself says it does not claim. Grouped by the page it would feed in the proposed site plan.

As at 16 November 2026, newsroom.sgit.ai v0.14.124, read through its markdown twins. Five read-only passes; each article skimmed from its abstract, its short list and its “what this does not claim” section. A row is a pointer, not a reprint.

Fit is a sorting label for this map, not a judgement of an article: direct is the ABP itself, supporting is evidence, buyers and selling, adjacent is useful context. The 77 items judged not for this site are left off.

In action · 11

Web Summit 2026

Article, dateThe ideaFit · becomesCautions
Section: Working with agents
fetched 2026-11-16
Agent teams in practice, newest first: Web Summit (around 250 named agents), the inbox team as a recipe, agent@riskmandate.ai 178 emails/47 decisions, re-anchoring, a second reader.supporting
article-reprint-link
Web Summit agent counts ('around 250') differ by article; recount before quoting.
PrompTugal.news: the agent newsroom at Web Summit
2026-11-15
Desks of agents captured, mapped, wrote and reviewed 73 signed articles in ten days, with an accountant counting costs and an outreach agent whose fifty drafts none could send, because only a person sends.supporting
example
Figures from repository on 15 November; counts from commit subjects; dollars API-equivalent; 'about 53' articles is a byline count (73 vs 53 needs reconciling).
The team's operating system: web-summit.sgit.ai
2026-11-14
Six people and their agents ran Web Summit from one git-built site with four layers: people, agents with keys and inbox lanes, work as files, and the outside world including the RiskMandate X-ray.supporting
case-study
Figures from repository on 14 November including post-summit work; dollar figures API-equivalent; keys per device, NFC exchange and signing as a graph were proposals only partly built.
Ten days at Web Summit: the map
2026-11-14
Six people and around 250 agents built a team operating system, newsroom, storm site, fringe guide, a rating standard and native apps in ten days; this is the map of the series.supporting
section
'Around 250 agents' counts distinct agent names, including renamed ones; 'more than 60 at once' an upper bound; commit total approximate; dollars API-equivalent. Mentions 'a draft standard that rated twenty companies', which needs reconciling with the no-score-on-an-ABP rule.
Web Summit team part five: last day of setup
2026-11-04
On the last setup day the team gave every person a key, added agents per person including a token accountant and synthetic users, and its behaviour policy agent again caught its own agents' mistakes.supporting
case-study
Numbers are the team's own, not an independent count; synthetic users tested designs, not real people.
Web Summit team: agents run alone, ABPs catch our own
2026-11-02
On a day of agents running on schedules, the team's own behaviour policies caught a procurement agent writing outside its folder and a session signing commits under another agent's name.supporting
case-study
Counts from the repository on 2 November with the day still going; some of what is on the site is still a plan; five of nine people not started.
Web Summit: humans and their agents as one team
2026-10-31
The first team with many humans needs clear roles, personal CRMs synced by shared ids, hour-level capture-to-action, and briefs suited to people with little time, worked backwards from the summit.supporting
section
A brief; decisions pending at time of writing.
Web Summit team brief: an event site and company map
2026-10-30
Brief to build the Web Summit team site from a template proven on three teams, centred on a public-material map of companies attending, with keys per email agent.supporting
section
Brief at time of writing; company map from public material only.
The ultimate insider: a talk built by nine agents in a day
2026-10-29
The hack::soho talk site was built in one day by nine agents and an editor, each with a role, ABP, session key and slot, with one writer per file enforced by the gate.supporting
case-study
Web Summit 2026: the team plan
2026-10-26
A briefing for five people at Web Summit: usage then the sale, five-minute onboarding, intelligence, outreach and content tracks, run from a summit newsroom.supporting
section
The Web Summit newsroom does not exist yet (at time of writing). Intelligence from public material only.
Native apps by agents in five days
2026-11-15
Agents built four iOS apps from one workspace, including a RiskMandate app, and an Android stand app with a signed build four hours after first commit.adjacent
none
Figures from repositories on 15 November; commit counts approximate; Gemma speed from one run.
In action · 16

Case studies (new)

Article, dateThe ideaFit · becomesCautions
Our workflow on Odoo 20, built and measured
2026-11-16
On the official Odoo 20 image, a credential-less bills agent's five proposals went through a policy check, signed approval and narrow broker, and the broker's grant was measured from Odoo itself into ABP rows.direct
case-study
Approval simulated, no person approved; Odoo's own agents and MCP not run (Enterprise-only, from documentation); no email delivered; extraction rule-based by design. Found that system-of-record CRUD is not yet among the ABP grammar's published capabilities, and a create-only grant is impossible in Odoo.
A teacher and agents she can trust
2026-11-03
A small business owner will only keep agents she can trust, so every agent gets an ABP saying what it is for, what it reaches and what holds each line: they draft, she decides.direct
case-study
Teacher fictional (invented); flows are designs, not set up for a real teacher; some barriers are Expectations (marked); prices read 3 November 2026.
The grant, not the prompt: mandate versus reach
2026-11-02
A user expected Claude to read his email only when asked, yet an unrelated chat listed his inbox because authority was granted at the connector, not the prompt: the mandate-reach gap RiskMandate measures.direct
case-study
One user's setup seen on screen on 2 November 2026, account not audited; vendor documentation (read-only) and product (draft, label, send tools) disagree, reported unresolved; Gmail scopes undocumented.
RM-CHAT-ONLY: an agent that may only chat
2026-10-23
The smallest useful mandate, chat only, built as a line, skill, CLAUDE.md block and fail-closed hook, and tested: the line alone falls under a fallback drill while the hook holds.direct
case-study
Run on the lead's own accounts. Found seven things to fix in RiskMandate's own policies, including hooks that fail open and a line that does not say who can lift it.
Inside the Pellbridge agents: six ABPs replayed
2026-10-20
An invented bank's six agents each have an ABP with the FCA's four elements filled in, and one replayed day shows every incident came through a gap row while rows kept by a boundary held, with breaches falling from 16 with no policy to none with rules kept outside the agent.direct
case-study
Everything invented (bank, people, agents, customers, systems, numbers); FCA has not endorsed it; outcomes scripted and levels modelled; weights illustrative; identity, kept_by, holder, decision and execution fields are proposals not in the ABP today.
Allowed is not right: answering Northstar
2026-10-18
An ABP answers SECURE (what an agent may do, enforced outside the model) but cannot answer VERIFY, which needs a decision policy saying what the right outcome is; a twin shows refusing is not the same as being right.direct
case-study
Twin is a simulation, synthetic orders, scripted mistakes, no model called; minutes are assumptions; twelve of thirteen cases invented; the mapping to SECURE/OBSERVE/RESPOND/VERIFY is the site's reading.
First ABP-governed SGit page merged into Ruflo
2026-10-17
Mondweep Chakravorty's documentation page for an ABP compiled into Ruflo's rules and gating SGit checkpoints was merged into ruvnet/ruflo, the first commit there to mention SGit or an ABP, and it states its own limits.direct
case-study
Documentation only, Ruflo runtime unchanged; no formal approving review or checks recorded; not an endorsement of every detail; 'first' is from commit history only; code read, not run; the runnable example PR #1 was still unreviewed.
Agent loses its workspace: SGit x Ruflo proof
2026-10-17
In a 104.7-second replayed run, an ABP-governed adapter saved Ruflo's memory as an encrypted SGit checkpoint, the project was deleted, a fresh process restored it with a matching SHA-256, and three safeguards held: out-of-scope refused, unapproved deletion stopped, tampered copy rejected.direct
case-study
Proof of concept, not a product; recall by meaning not validated; facts synthetic (Kestrel a test name); analysis by an AI reviewed by the site's agent; code read, not run, by this site.
An agent's logs are a self-report
2026-10-16
Answering a GRC adviser, this site's agent is the worked example: evidence written outside the agent is more honest than its account, authorisation limits must be enforced outside the model, and owners should be named per mandate and the ABP published.direct
case-study
One agent is not a sample; the policy and hook are not independent of the agent (it could edit them); no outside red team has tested it; research numbers are setup-specific; the post's poll and case-study numbers could not be traced and are not used.
Moving an agent's rules from hope into hooks
2026-10-14
Of one agent's eleven rules, only one was kept by something other than the agent; moving them into hooks in a sandbox held some fully and others not at all, and only works if the agent cannot edit the hooks.direct
case-study
Tested in a sandbox; the generated hook pack is a proposal: 'The generator, the tests on every change and the gateway compilation are not built.'
A second reader gate on every draft
2026-10-09
Two emails went out in the lead's voice despite a written rule because the agent was the only enforcer; a hook making a second model approve every draft, failing closed, turned hope into a boundary.direct
case-study
The article's own librarian flags that its cost figure does not reconcile and is not reproducible from the stated figures.
The Mandate Stack: eight layers in production
2026-10-06
RiskMandate runs on about fifteen agents and one human in eight layers, one written mandate per agent, everything a graph, the human the only one who sends.direct
case-study
Agent count needs a recount: this article keeps 'about fifteen' from the CRM agent's briefing while the-agent-team-as-it-runs names twelve; the article notes the difference itself.
The agent team as it runs: twelve agents
2026-10-06
Twelve agents on dedicated accounts, each with a brief and an ABP, encrypted vaults as the only memory, messages as files, and a mailbox nobody sends from, with barriers typed honestly.direct
case-study
Counts twelve agents where the-mandate-stack says about fifteen; it has a 'What is not working yet' section; infographic carries no live data.
Six agents, one inbox: access policies graded
2026-09-29
Running six agent roles on dedicated accounts showed the account is the blast radius and an access policy is only as real as its worst-enforced row.direct
case-study
Finding three: the policy written from documentation was wrong; reach differs between Cowork and Claude Code on the same seat.
The cost of your agents: a partner integration
2026-11-03
A cost tool sees what each agent spent but not what it is for; joined to a behaviour policy with a cost row, spend can be judged agent by agent against its mandate.supporting
case-study
Tallyline, Pellbridge Bank and the month of usage are invented; integration is a design, neither product connected; savings assume findings work. Article says each agent gets a 'verdict against its mandate', which sits awkwardly with the no-verdict rule for ABPs.
The RiskMandate inbox team as a copyable recipe
2026-10-28
The RiskMandate email workflow, written by its own agent as a guide: mailroom, twelve agents each with a mandate and ABP, file-based CRM, Gmail rules where no agent sends, a security agent first and last.supporting
case-study
Notes Google Calendar is where Gmail's draft model breaks.
In action · 8

Vendor-shaped examples (new)

Article, dateThe ideaFit · becomesCautions
Odoo agents can now write to the system of record
2026-11-16
Odoo 20 lets agents create and update records unattended, bringing the assist-to-execute line to small business; RiskMandate would write an ABP per Odoo agent with the grant measured from Odoo, ideally by a module.direct
page
A strategy and proof of concept, not a product, nothing sold; proof used Odoo Community images, Odoo 20 and Enterprise agents not tested; policy in the figure illustrative; corrected 16 November from a separate agent's build.
Who else controls AI agents at Web Summit
2026-11-05
Eleven agent-control companies, read in their own words, mostly work while or after an agent acts; RiskMandate works before, on the reach-mandate gap, so most are barriers an ABP can name and partners as much as competitors.direct
page
Placements and 'where we differ' lines are RiskMandate's view; some attendance from a third-party list, unconfirmed; a few details on the page need small corrections.
WSO2 Agent Manager and where behaviour policies plug in
2026-11-03
An open agent control plane supplies identity, gateway, runtime and traces; behaviour policies add the mandate on top, at the gateway, as a SHarness isolation tier and in the same trace.direct
section
The MVP is a design, not built or run; WSO2 not consulted, read from its public pages on 3 November 2026; GitHub 1.0.0 tag dated after the GA date in the press release.
An analyst's blueprint for RiskMandate
2026-10-21
An analyst's note says control what agents do, not just what they can access, and lists vendors moving into the space; RiskMandate's answer is not to intercept tool calls itself but to supply the context-fitted policy above every enforcement point (prompt, hook, gateway).direct
section
Note AI-assisted and shared by a power user; vendor capabilities are their own claims read 21 Oct 2026, not tested; RiskMandate supplies policy, not a product in each layer; evidence trail is a demo (Pellbridge invented); names third-party vendors, keep to facts and dates.
Resilience when the workforce includes agents
2026-11-04
A third-party resilience rating covers operational, data, financial and people disciplines but does not yet count agents in the workforce, which is where RiskMandate, sgit and SHarness meet it.supporting
section
OCRR scale and scoring not published, ratings marked illustrative; vendor products not tested; intersection is the author's reading, not a partnership; SHarness whole-agent launcher still a design.
From Lovable to your own newsroom: a brief
2026-10-31
A brief for a collaborator to move his agent-reach apps from Lovable to GitHub Pages, his own domain and a newsroom, for RiskMandate to amplify; reach tools as his business.supporting
none
A brief for a collaborator; outcomes not yet achieved.
Agentics ready to sell: market, path, marketplace
2026-10-16
After Agentic Engineering Live, governed agent-hive engagements whose runs leave a vault and ledger the client keeps are ready to sell into a market thin on evidence, with a consultant marketplace to deliver them.supporting
section
No paid engagement yet; the runs were not policy-gated (enforcement shown in a separate proof of concept); market map a judgement; Gartner figures from press releases; the marketplace is a design, nothing built.
Starting AgentiCCells
2026-11-15
A new company sets up small open agentic teams in the customer's own environment, with ABPs and a control centre, keeping no state on its side; research found nobody packages these together as a service.adjacent
none
Prices a first proposal; four decisions open; company a day old, not yet a service anyone has bought; vendor facts compiled by a research tool.
In action · 3

Invented twins, labelled

Article, dateThe ideaFit · becomesCautions
Review of Hack the Agent and ReachCheck
2026-10-31
Two Lovable apps built at the Vibathon ask what an agent can reach and what stops it: a prompt-injection game and a browser-only reach review tool; reviewed with invented input only.supporting
example
Used with invented input only; the five game agents are fictional. ReachCheck flags red, yellow or green, a traffic light that riskmandate.ai's no-score rule forbids on a behaviour policy. Built by a collaborator; not a RiskMandate product.
An agentic console: a simulated control room
2026-11-15
A team of agentic teams is sellable only with a console to see and control it; a simulated one draws accounts as trust zones, rewinds time, shows signed messages and stops that say when they land.adjacent
none
A simulation: runs, messages, commit counts, boards, models and key fingerprints are made up; nothing live, stop buttons stop nothing; customer history is a reading of own ventures, not contracts.
Move the cloud, not the code: twins
2026-10-12
Workloads stay on hyperscalers because moving means changing code; an API twin lets the code not know it moved, and the twin can carry behaviour policies the calling agents cannot skip.adjacent
none
The S3 twin is 'the only finished twin' and its API surface 'still needs confirming'; IAM is partial; EC2 not finished.
In action · 27

Example ABPs and the library

Article, dateThe ideaFit · becomesCautions
RiskMandate and Lovable: an ABP for an app builder
2026-11-09
Lovable's agent can change production and its controls are real but default open; an ABP records what each project's agent was asked to do, which settings should be 'Ask each time', and who signed it off.direct
example
Not run on Lovable; mapping from its documentation read 9 November 2026; no relationship with Lovable; platform AIUC-1 certification covers the platform, not the project (reprint must not borrow the conformity word).
RiskMandate and Cloudflare: room for a signed mandate
2026-11-09
Cloudflare runs and controls agents with open defaults, and its own writing says a scope states what an agent may do, not who answers when it is wrong: the record an ABP provides.direct
example
Not run on Cloudflare; documentation read 9 November 2026; WriteGuard and identity-aware AI Gateway in beta; the 'triage, close none' mandate is RiskMandate's invented example on Cloudflare's 'Joe'.
RiskMandate and Manus: an ABP for a general agent
2026-11-09
Manus has wide reach and real controls whose documentation names what they leave open; an ABP lists reach, mandate, gap and the control holding each row, living as read, set, gate and attest.direct
example
Not run on Manus; deployment invented; 'gate' and 'attest' are proposals; no relationship with Manus.
One writer, one key: a register of writer agents
2026-10-31
After two writer agents collided, a register lets only one writer registration be open, each commit signed over its files, and the validator refuses six cases.direct
example
Remaining weakness stated: the lead's approval is still a note in the same repository.
Rules of the game: which files an agent may change
2026-10-22
Agents that can change their own checks can pass any release, so a five-zone policy and what GitHub can enforce on a public repository define the barriers.direct
example
Nothing here is enforced yet. The policy is proposed; no hook, guard or token change is in place.
A skill behaviour policy, measured on 899 skills
2026-10-14
A skill behaviour policy compares what a skill says it does, what it can reach and what it is given, then generates the control that closes the gap.direct
section
Read statically (not run); the article calls the '1 in 20' video claim an overstatement of the study.
System prompts rewritten as behaviour policy rows
2026-10-14
Claude's published system prompt mapped into 82 behaviour policy rows shows many rules could be held by the harness; both prompts examined lack a mandate, an authority order and what keeps each rule.direct
example
The ChatGPT prompt is leaked and 'unverified'; the JSON ontology is a proposal.
A behaviour policy for every agent skill
2026-10-14
Sean Turner's natural-language CI control for agent skills is a behaviour policy for a skill (mandate: what it is for; reach: what it touches; gap: the risk), and natural language alone needs deterministic rules generated from it.direct
example
Sean's tool described from a conversation, code not seen or published; the two skills in the figure are illustrations; licence to operate and insurability are designs, not products a customer runs today.
Hope or enforcement: one agent, three designs
2026-10-08
One customer service mandate built three ways shows every design has rules for everything; the difference is how many are hope (kept by the model) versus boundaries outside its reach.direct
example
The shop is fictional. Hope percentages are counts of rules by barrier type; present them as counts, not as a rating of the ABP (rule 1).
Behaviour policy in practice (collection)
2026-10-08
RiskMandate's ABP applied to real agents in eleven articles: one inbox, a team of twelve, the 2026 personal agents, footprint and blast radius afterwards; the newsroom's own desk roles run on written behaviour policies.direct
section
Collection last updated 2026-10-08 (page footer sgit.ai v0.7.87); later ABP pieces (insurance series 28 Oct, the grant not the prompt, SHarness, Web Summit) are not in it. Its intro says 'the policy' alone and 'graded'; reword on reprint (rules 1 and 3).
Give the connector a twin before the agent
2026-09-24
Before an agent gets a Gmail or Calendar connector, a journal of every call (a twin) turns 'we cannot undo it' into a named list the mandate can be written against.direct
section
The replayed session is invented (invented people, .example addresses); the Connector Twin is published as a business plan 'for somebody to build', not a running product.
Provenance is the product
fetched 2026-11-16
Design: a newsroom that shows its walkable chain (who reported, checked, what it cost) earns trust a black box cannot; every claim links to evidence.supporting
homepage-block
Page says it describes a design; the £8.40 story is illustrative, not a live ledger. No .md twin.
Synthetic users for your website
2026-11-04
Invented personas, each run by a sub-agent seeing only screenshots of a real browser, show what a site looks like to each buyer; run three times including on riskmandate.ai.supporting
section
Not user research; nothing a persona says is evidence about a real person; personas are invented.
English with Ines: a teacher's site as vault app
2026-11-03
A worked vault app for a teacher's site: lessons, spaced practice, a tutor bound to one lesson, and seven agents each with an ABP and the barrier that holds each line.supporting
example
Teacher, students and pipeline are fictional (invented); live tutor not tested; progress per browser only.
Sources and their shadows
2026-10-25
Every reference carries primitives worth recording (licence, access, links, quotes, quality, history, payee); an unreadable source can be seen through its shadow of citations, disagreements between shadow and source are evidence, and sources used should be paid, more when they certify the use.supporting
page
Its own 'does not claim': licences not yet recorded; a shadow is not the article; 'paywalled' is the watch's classification on 23 Oct; the payment funnel is a design, no source paid; 'who got it right' is a record to build, not a verdict.
Levente Simon's governed development method, read here
2026-10-21
Levente Simon's method for governed agent teams arrives at many of the same ideas (a session cannot write what enforces its limits; rule text is a guard-rail, not a boundary), and by its lowest profile the newsroom falls short on two of eight layers, one because the publishing agent can change its own checks.supporting
article-reprint-link
One reading of v0.7 of the method on 21 Oct 2026; self-assessment is the site's reading, not an audit; settings differ; suggestions only.
The AI Conference 2026 read for agent governance
2026-10-16
Reading the programme and 35 slide decks, the loudest shared idea was that policy must be enforced outside the model, silent failure was the commonest story, three of six reported items trace to the stage, and the talks disagree on five questions.supporting
section
No transcripts; not found is not false; the 43% poll, 1,200 developers and 1,500 agents could not be traced; vendor numbers are vendor numbers; the readings are the site's.
Four AI risk frameworks against one agent
2026-10-13
A paper comparing OECD, NIST AI RMF, Japan's guidelines and the EU AI Act classifies systems before deployment; for agents the unit should be one agent in one deployment, described by its ABP.supporting
section
Rule 7: the article quotes NIST subcategory wording in a mapping table; any reuse must keep standards to titles only (EU AI Act may be quoted). Avoid 'aligned' language.
The records ran out: evidence labels at the edges
2026-10-12
A reported case of a fresh agent finding a shared contract from build records shows where records stop: label every fact by how it is known, and treat a skipped test as a control that did not run.supporting
section
Reads a third party's published account; nothing was tested.
How to run synthetic users
2026-10-10
Five to ten invented users drive a real browser, are interviewed and their findings ordered by cost; run on store.sgit.ai, riskmandate.ai and the newsroom.supporting
example
All personas invented; study vault linked from the article.
A Mac of the agent's own: a business plan
2026-10-09
A business plan for per-run agent Mac desktops built from vaults within Apple's licence, justified by three behaviour policies for one agent showing unbounded excess shrink.supporting
example
A business plan 'for somebody else to build'; not running.
An open AI baseline control framework, graphed
2026-10-07
A CC BY-SA framework of twenty AI controls, including an Access category the mapped standards lack, was turned the same day into a graph linking nine controls to ABP terms.supporting
section
Mapped to NIST AI RMF and ISO/IEC 42001: cite titles only, never their text (rule 7); avoid any 'aligned' language.
2026 personal agents read through behaviour policy
2026-10-06
The 2026 personal agents (Muse, Spark, Autopilot, Alexa+, Siri, OpenClaw and others) differ most on the barrier and on who can read memory; a design on vaults, an enclave and the browser keeps keys from every vendor.supporting
article-reprint-link
Reads named vendors' products from their pages: facts and dates only. The enclave part of the design does not exist yet per the article's own 'What exists' list.
Desk: Vaults & provenance
fetched 2026-11-16
For anyone who needs evidence they can open: encrypted vaults, keys, graphs and the publishing method; includes 'the proof is two clicks behind the claim'.adjacent
article-reprint-link
No .md twin.
Storm Charles: a newsroom built in a night
2026-11-11
About thirty agents built a storm site overnight with a public weather face over a data room where every figure carries its source, read time and a fact, evidence or hypothesis label.adjacent
none
Not a warning service; figures are the site's own on 11 November and move; site a night old, no agent signs with its own key.
One Human, Nine Agents, Nineteen Hours: RC2 draft
2026-11-03
The second draft of the book of the talk shows the agent factory's controls on the page: reviewers catching their own team, an objection kept, and a hold sixteen minutes before the talk.adjacent
article-reprint-link
RC2 is a draft; chapters 4 and 5 had not passed every check; scenes dramatised; test readers are fictional personas.
Knowledge that maintains itself: semantic graphs and vaults
2026-10-14
Turn an organisation's policies and internal knowledge into a fractal semantic graph that agents maintain, brief people and agents from the graph rather than the documents, and carry the brief in an sgit vault.adjacent
none
Tokens estimated as characters divided by four, at newsroom v0.12.39.
In action · 4

Try it

Article, dateThe ideaFit · becomesCautions
Inside the Northstar refund twin
2026-10-18
A tab-by-tab tour of the Northstar refund twin: thirteen synthetic refunds through four control designs and ten controls, every decision replayable, with no model called.direct
example
Synthetic throughout (only Northstar and its $120, $150 and $200 come from Ashish Rajan's article); mistakes scripted; minutes are assumptions (six a case, three an approval); one process.
Replicating the agentic inbox, phase by phase
2026-10-02
A walkthrough to copy a multi-agent email setup: own accounts, one session with a behaviour policy written before the first run, then roles talking in files, with agents never pressing send.direct
page
The four-column ABP picture is a mock-up with fictitious data.
From spreadsheet CRM to a folder per person
2026-10-12
Move a spreadsheet CRM to one Drive folder per person, worked by a Claude Project whose instructions are a six-line behaviour policy, before any vault is needed.supporting
example
Becoming an official publication
2026-10-30
After withdrawing an article that repeated unnamed-source allegations about named organisations, the newsroom adopts a publisher, editor, standards and corrections log.adjacent
faq
In action · 1

What runs, what is designed (new)

Article, dateThe ideaFit · becomesCautions
What runs, what is designed
2026-10-10
Dated authority on what is live on the newsroom vs designed vs stopped; e.g. 434 articles, 6 roles with written behaviour policies, reading meter live; selling the graph, paying the fact creator, trust as a service: designed, not built.supporting
page
Dated 10 October 2026; stale by five weeks. No .md twin.
Product · 33

Where a rule is kept (new)

Article, dateThe ideaFit · becomesCautions
The email agentic cell: one gateway touches Gmail
2026-11-16
Usual Gmail connectors give every agent everything with no enforcement; instead one gateway agent alone holds the token, others work on a vault twin and propose signed actions, with behaviour policies at five altitudes.direct
example
A brief and architecture; none of it built; Google verification requirements need confirming; signed reads record a key opened a message, not understanding.
CI behaviour policies: a pipeline's reach exceeds its job
2026-11-16
A pipeline that makes one repository and one DNS record per request holds power over every repository and any subdomain; a behaviour policy for the pipeline names reach, undo, limit and a boundary outside the grant.direct
example
Nothing exploited, a reading of grants on 16 November 2026; reviewer and branch protection not checked; proposed limits and brokers do not exist (FAC-007 proposed); capability names follow the abp.sgit.ai grammar but are not in the published list. Own system, so entitled to read.
Coding standards are behaviour policies
2026-11-13
A coding standard describes behaviour as intent, so read as a behaviour policy the 31 house rules kept only by habit are a mandate with no barrier, and the policy makes that gap visible.direct
section
Cart example invented; cross-language, cross-team mapping is a proposal; rule figures from coding.sgit.ai measurement of 25 August 2026.
False positives and the layered harness
2026-11-02
A behaviour policy layered from network to business-logic fields should refuse less of what is right and more of what is wrong, and learns as gaps found in testing become rows.direct
faq
No measured false positive rate; most of SHarness is proposed (behaviour rows, doors, layering, CI pipeline are designs, not running code); order example illustrative.
Introducing SHarness: the ABP, executed
2026-10-29
SHarness turns ABPs into Process Behaviour Policies enforced by the kernel or a separate service, so Settings and Expectations become Boundaries an insurer can price; released as sharness-ai 0.2.0.direct
page
Most of SHarness is designed, not built; nothing binds the agent itself yet; no protection against root; measurements are the project's own; translations not run against targets; security reviews by the project's own agent; the cloud service is a proposal; 'integration point' is not a partnership.
Proven controls: who holds each control
2026-10-29
Policy must leave the PDF and become runnable and testable, and the missing column is who holds each control: only a Boundary held above the agent can be proven not switched off; six controls tried to break.direct
page
Reads Hem Pant's article and Sondera's paper as published; not their view of RiskMandate. The ABP's move to the Open Security Summit is a proposal in progress.
Process behaviour policies: the most locked-down Linux process
2026-10-24
A policy derived from one traced run, enforced by the kernel with seven walls, stopped a root probe at 92 to 187 ms a run; the policy's size measures the software.direct
example
Measured in the newsroom's own container only.
A sidecar repository the agent cannot write to
2026-10-22
A separate repository holding ABPs, approved hashes and keys verifies each release against policy and signed attestations before deploying, assuming a fully compromised agent.direct
section
Nothing here is built. It is a design, and a proposal to the owner.
57 checks taken apart into 8 behaviours
2026-10-21
Every test checks for a behaviour: the newsroom's 57 release checks reduce to eight behaviours, CI blocks a deploy for 19, and in ABP terms only one is a boundary out of the agent's reach because the agent can change the files that check it.direct
example
57 is this site's count today; grouping is the writer agent's; eighteen test files (1,061 checks) grouped as twelve rows; proposals not yet applied.
The ABP guards the door; the CLI uses the window
2026-10-17
In the SGit x Ruflo run the ABP compiled into Ruflo rules is enforced on the adapter's tools and every guarded step behaved, but Ruflo's own command line reaches the same memory file around the gate: a policy holds only on paths through its enforcer.direct
example
Not a vulnerability report, nobody tried to bypass it; the author documents the path; fixes are ideas (native plugin registration is his follow-up); code read, not run.
Cedar and the policy vault
2026-10-14
Adopt Cedar rather than compete: permit maps to the grant and forbid to the barrier, but Cedar lacks the why, the version in force and the evidence, which a vault carries.direct
section
The MVP is a brief; Cedar skips a policy that errors, so an invalid forbid fails open.
Agent hooks across ten platforms, October 2026
2026-10-14
Hooks before a tool call are now standard across at least ten agent tools, but they differ on the security properties that matter (organisation-only hooks, fail-closed, trust for repository hooks, model-output visibility) and every vendor calls them guardrails, not walls.direct
page
Map read from vendor documentation on 14 Oct 2026, partly by research agents; cells marked unknown not verified; one test in one tool in a throwaway repository, not a measure of any product's security; failure modes described, not demonstrated.
Policy as code needs an ABP first
2026-10-14
Policy as code answers only 'is this request allowed'; an ABP holds reach, mandate and reason, gap, rules no engine can see, owner and who can change each rule's code, and it can be compiled to Rego or reverse engineered from existing code.direct
page
The round trip is a demonstration (five rows, short generator, seven requests), not a product; session table is one agent's, counted 14 Oct 2026; GitHub scope not tested by trying.
A behaviour policy beside every prompt
2026-10-13
Every agent should take an ABP, written to a published standard, as a third input beside system prompt and prompt, which the vendor accepts, enforces where it can and answers for, as aviation and payments learned.direct
page
Labelled a proposal; 'published standard' does not yet exist, so no conformity language.
Microsoft MXC/ACS and the shared responsibility line
2026-10-11
Platform controls such as Microsoft's MXC and ACS bound what they can see (paths, addresses, tool calls); the behaviour policy says which rows become boundaries and the business logic above stays the customer's.direct
section
Named third party: facts and dates only (MXC GA 7 October 2026); the article's 'most complete platform control' wording is evaluative and should not be carried over.
Re-anchoring rules through every summary
2026-10-10
Rules given only in conversation are lost when a session is summarised; keeping them in a behaviour policy file the harness prints back after each summary, with a canary report, restores them.direct
section
A permission prompt that kept the reason
2026-10-07
An agent permission prompt asked for a decision without saying who asked, why or what reach would change; the proposal is a 'why card' that becomes a risk acceptance as risk rises.direct
section
The why card is a proposal.
Footprint and blast radius added to the ABP
2026-10-02
Proposes adding footprint (what the agent actually did, read afterwards) and blast radius (what a reach row would cost today) to reach, mandate, gap and barriers.direct
section
Explicitly 'a proposal to extend the vocabulary' of the ABP; the gap map is for an illustrative inbox agent.
Desk: Building with agents
fetched 2026-11-16
For engineers running agents day to day: the agentic inbox, agents off the laptop, a second reader as a gate on drafts, re-anchoring rules; series 'The agent team, at three depths'.supporting
article-reprint-link
No .md twin.
Lovable with SHarness, and the tampered extension
2026-10-31
A tampered editor extension shows agent security is end to end; SHarness cannot reach inside Lovable's cloud but can contain the local clone and agent, and its policy rows would refuse each step of the exploit.supporting
example
SHarness runs only on Linux; the tampered build targeted macOS and Windows; the launcher that would bind an editor or agent is not built; only two relevant rows run today. Facts on Cursor's incident from its own notice.
The push is the authorisation for an agent key
2026-10-28
Pushing a session's public key to the team repository counts as authorising it, moving the key decision into an existing publishing boundary; role and session are separated.supporting
example
Security card conductor-0017 on who can designate a recipient was still open.
What the bitcoin bounty proves about php2shell
2026-10-26
Checked against primary pages, the bounty took three hours once reachable, every vendor's detection has a gap it names, all published hardening is partial, and patching is the only complete fix.supporting
example
Read from vendors' and the operator's own pages; nothing tested.
A security rule applied literally made code less secure
2026-10-25
A sound same-origin script rule led agents to copy shared code into six repositories; a person caught it and replaced copies with one versioned, hashed source.supporting
example
When the vault is used in SGit x Ruflo
2026-10-17
In the twelve-step SGit x Ruflo run the encrypted vault is written in step 5, read in 8 and rejected in 11, on one machine inside the adapter, which is the right shape for a recovery copy but does not yet show the vault leaving the machine.supporting
example
From the screen and the code; inferred arrows dashed; not a review of SGit's cryptography (tamper caught by the adapter's digest check).
The tokens hook: a token behaviour policy
2026-10-15
A small observe-mode behaviour policy about tokens, built as a hook, shows what each session reads, caches, writes and costs, which a prompt cannot do because the model cannot see its own usage.supporting
example
One tool, one version (Claude Code 2.1.300); prompt test small; costs at list prices; other tools from documentation; tokens.hook.sgit.ai does not exist yet, it is the proposal.
Hooks are kernel code: a maturity model
2026-10-14
Agent hooks are small code with kernel reach; a five-level maturity model maps how a rule is kept, from prompt to off the machine, with the first ten hooks mapped.supporting
section
The hooks site and feed is a brief, not built. The maturity levels rate hooks, not an ABP.
How cheap is it to undo?
2026-10-13
How much an agent can ship unaided depends on how cheap the output is to undo and how mature the controls are: the newsroom releases itself, email still waits for a person.supporting
section
How agents decide, and when they stop (collection)
2026-10-08
Six articles making one argument: deciding needs more than yes/no, and an agent whose range is the point needs its stop designed in, tested by one agent built three ways with an ABP each.supporting
section
Collection last updated 2026-10-08 (page footer sgit.ai v0.7.87); later ABP pieces (insurance series 28 Oct, the grant not the prompt, SHarness, Web Summit) are not in it. 'Agency is not a yes' has a seven-level scale for deciders; on riskmandate.ai keep it about decisions, never a level for an ABP (rule 1).
A locked-down agent desktop is hard to rent
2026-10-07
Nine properties a safe agent desktop needs can each be bought somewhere but no product offers all; the proposal is one mandate per task and a fresh desktop configured from it.supporting
section
The build is a proposal; vendor prices are from their pages at the time and need dating.
An inbox only your agents can write to
2026-11-04
A conductor publishes one append lane per agent with each token sealed to that agent's public key, giving a private channel per agent built entirely from public material.adjacent
none
Pattern one evening old; self-tested, no external review.
A key for every person on the team
2026-11-04
Each person's conductor issues them receive and sign keys, so sealed content can sit on a public site and agents treat only verified signed messages as instruction and everything else as material.adjacent
none
A working short-term design; one security review, no external audit; two open findings, so nothing truly sensitive is sealed to a person key yet.
PKI everywhere: signatures and an agent's first key
2026-10-20
Every message, secret and handover should start with a signature, and an agent gets its first key either by making its own confirmed by a person or by an identity key issued centrally after approval; identity is not permission, which is the ABP's job.adjacent
none
Pattern B is proposed, only Pattern A built; register has one agent key and no confirmed links; platform support as found on 20 Oct 2026.
Handing secrets to agents and browsers
2026-10-15
SG/Send should focus on handing a secret to an agent or browser: the recipient makes its own key pair, the owner encrypts and signs to it, the drop is read once, and the secret goes into the tool, never the chat.adjacent
none
Ran locally with one dummy secret; refusals are observation not measurement; first secret not solved; prior art from documentation; SealDrop a working name; the friendly code can be recovered offline from the public id.
Product · 11

What an ABP is

Article, dateThe ideaFit · becomesCautions
Ten agent risks in Anthropic's IPO filing, mapped to ABPs
2026-10-27
Quotes ten agent risks from the 26 October 2026 Nasdaq registration statement with page numbers and maps each to what ABPs do: write down what an agent may do, keep it out of reach, prove it with outside evidence.direct
page
Some answers built and measured, some proposed; each says which. The filing is preliminary and its wording may change.
Application behaviour policies after php2shell
2026-10-22
The ABP method applies to an application under remote code execution: write its reach, purpose, gap, barrier and blast radius per row, so bringing a site back is a decision about known rows.direct
page
The worked council website example is invented; RiskMandate's own sites are the measured contrast.
The delegation record: one chain for one workflow
2026-10-22
The next step is one delivered chain for one workflow: business requirement, approved delegation, control mapping, evidence, remaining gaps and invalidating conditions, starting with our own inbox.direct
page
Plan stage; pilot zero is internal (own inbox). Notes Gmail has no draft-without-send permission.
Approved business mandates into tested controls
2026-10-21
Companies already have behaviour policies written for people (delegated authority, approval limits, four-eyes rules); RiskMandate's job is to turn these approved business mandates into ABPs and tested controls, say where a rule cannot be enforced yet, and keep them true as the deployment changes.direct
homepage-block
Review is AI-assisted (a ChatGPT session); rows and tables illustrative, no customer has a coverage explanation yet; nothing is a pilot result; decision and workflow policies explored, not deployed; accepts that the Cedar work is a demonstration, not a compiler, and Pellbridge explains, it does not measure.
Drift is the mandate moving
2026-10-21
Drift (policy, behavioural, goal, authority) needs something to drift from, and a versioned ABP is that baseline: three of the four drifts are measurable now by re-measuring the grant, checking actions against rows and checking what keeps each rule; goal drift is open.direct
page
The analyst's note is AI-assisted; Pellbridge is a demo, no customer runs it; re-measuring a grant is not continuous yet; goal drift not solved; procurement rows invented; OpenAI's account is its own.
Can a key leak without breaking the ABP?
2026-10-20
A key can leak without breaking a capability row because the key is content, so an ABP needs content rows with barriers at three levels: told (prompt), checked (a hook on every input and output) and never given (something outside the agent holds the key).direct
page
The hook is an example tested on seventeen cases and one session; experimental rules expect false positives; fake keys only; a hook the agent can change is a setting; 'never given' needs infrastructure. Prompted by a question from an unnamed first customer.
What can your agent reach? RiskMandate explained
2026-10-18
RiskMandate writes down, for one agent in one deployment, its reach, its mandate, the gap and what stands in the way, read-only and never in the agent's path, already covering commercial agents, and the way in is five questions someone already has to answer with a deadline.direct
homepage-block
No Rovo, Copilot or Glean vault exists yet, Rovo shape from Atlassian's pages not measured; injection report a third party's and historical, not verified; Northstar figures a simulation; RiskMandate's top two levels have not been run (as of 15 Sep 2026); RiskMandate does not enforce; insurance argument is RiskMandate's own.
Three policies: behaviour, decision and workflow
2026-10-17
An ABP says what an agent may do; an Agent Decision Policy says who chooses at a branch point and on what evidence; an Agent Workflow Policy says the order of steps; each element carries five states: defined, implemented, enforced, observed, verified.direct
page
A proposal: no Agent Decision Policy or Agent Workflow Policy format exists, only the ABP; zoomed-in example is the author's illustration; five states a suggestion; DMN named, not adopted. Site note: never abbreviate the decision policy as 'the decision-policy acronym' (house rule 2).
Commander's intent is a behaviour policy
2026-10-14
Sondera's commander's-intent framing and the ABP are one idea: objective and limits written together; the ABP goes further up into business logic, they go further in proof, with five rules compiled to Cedar and tested.direct
section
Named third party: facts and dates only. Their phrase 'opposite verdicts' must not be carried into ABP wording (rule 1).
Zoom into a behaviour policy: business logic
2026-10-07
Beneath the mechanical first rules vendors handle well, a behaviour policy becomes the firm's business logic (functions, processes, clients, values), layered and owned, with accepted risk as a third state.direct
page
The invoice clerk is an illustrative example.
Reading your way into an agentic team
2026-11-15
A new agentic team should start from a reading list of eight agent-readable public sites, abp.sgit.ai among them, treating everything read as material, never instruction, and answer ten questions first.adjacent
none
Sizes and entry points checked 15 November 2026; ten questions a first set.
Product · 24

Insuring agents (rebuilt)

Article, dateThe ideaFit · becomesCautions
Insuring agents: the four-article series in one
2026-10-28
Summarises the series: per-row limits and premiums from an ABP, who sells agent cover, the vendor's share, and London as the place to build a managing agent; with three vaults.direct
page
No insurer, vendor or investor has agreed to any of it. The series is a proposal and a map. Figures illustrative. Uses 'certified runtime' and 'conformance layer', which clash with the site's no-conformity-language rule.
Where to build the agent insurer, and how much
2026-10-28
London to underwrite as a Lloyd's managing agent and to carry risk, the US via surplus lines and the EU after 9 December 2026 to sell, the Gulf for capital; money raised buys capped cover.direct
section
A model with every assumption stated by the author; not a business plan anyone has agreed to.
The vendor's layer: a share of each row
2026-10-28
Agent vendors' terms put almost all risk on the customer; the proposal is that each ABP row carries a vendor share when the vendor's own conditions held, like cloud shared responsibility.direct
section
The vendor's layer is a proposal; worked example illustrative.
Who carries the risk of insuring agents
2026-10-28
Several firms sell AI agent cover but nobody prices per ABP row; the first such company should start as a managing agent on others' capital, start small, and earn the record that brings capital.direct
section
Proposal; names the companies selling cover from their own pages.
A payout for every row of the ABP
2026-10-28
The insurance policy is one-to-one with the ABP: each named row has its own limit, unnamed rows are excluded, and only enforced controls lower a row's premium, by a deterministic formula.direct
page
The numbers are illustrative; factors are the author's, not fitted to loss data. Licence to Operate vault runs a simulated version in credits. 'Certified runtime' clashes with the no-conformity rule.
AIUC's insurance blueprint reviewed and mapped to ABPs
2026-10-24
The AIUC-led blueprint says insurers need accountability, reach, change and control evidence but not how a company produces it day to day; ABPs and AIUC-1 as a graph are that layer, with gaps listed.direct
page
Paper's coverage tower figure is hypothetical and illustrative; article lists what the paper has that RiskMandate still misses.
Who maintains the PHP you run?
2026-10-23
Every version of an open-source platform a product depends on should be signed off by its maintainers, buyers should require it, and RiskMandate is turning that into a paid model: map, engineer, custom version, retained maintenance.direct
page
Article says services, signed-off versions, insurance/licence links and RiskMandate's options are all proposals; prices are anchors 'not published pricing (the public site lists £10 to £1,500)'; author has a commercial interest. Possible contradiction with pricing.html, which shows only L1-L4.
RiskMandate roadmap: a spectrum of trust
2026-10-20
RiskMandate's path with each customer runs policies, understand, risk acceptance, contain, manage and insure, each stage answering the next question a customer asks, with insurance as the end game and acceleration running through all of it.direct
page
A proposal still under discussion between the founders, not the agreed strategy; stages have no dates; only the first stage is under way; no insurer is signed up; counts are the customer site's own, read 20 Oct 2026. Differs from the site's current ladder (ABP, Licence to Operate, Insurability Index).
Three decisions mapped across behaviour, decision, workflow
2026-10-17
Mapping agent memory deletion, the newsroom's release and an illustrative refund through behaviour, decision and workflow with five states shows the gaps are often unwritten decisions, visible only because each element was named.direct
example
The refund is a simulation, no refunds made; states for the memory scenario assessed from code and docs, not run; leak-scan flags were in scratch repositories, nothing leaked.
Case law for agents: precedent into policy into code
2026-10-17
Like English common law, agents should escalate cases the decision policy does not settle, record the person's decision as precedent, follow it for like cases, distinguish unlike ones, and compile precedents that hold into policy then code.direct
section
A proposal, loop untested; not legal advice; the two cases are small.
Approve at altitude: narrow, never widen
2026-10-17
Decisions exist at every altitude from strategy to a single run; each level signs off once and a lower level may narrow its parent's envelope (money, blast radius, data, people, reversibility, duration) but never widen it, which is mostly a mechanical CI check.direct
section
Envelopes can be gamed by splitting, need aggregate limits; check only as good as the numbers at the top; model flags are suggestions; design, not a product.
The policy draws the liability line
2026-10-13
The agent accountability 'shell game' works because nothing is written down; a behaviour policy's explicit prohibitions move where liability falls between user and vendor.direct
section
Answers a named person's post: quote and date, no evaluative wording.
Every risk is already accepted: by whom, until when
2026-09-24
A risk exists once the exposure does, so someone is already carrying it; the only choices are accept for a stated interval, fund the work, or fix it, and silence escalates.direct
page
Worked example is an invented company; the article labels a vault per material risk as "a proposal, not yet a practice". It also says taking risk acceptance to market via RiskMandate was hard.
Trust as a service (economics)
fetched 2026-11-16
Design for a payable fact-certifier role with warranted output and credibility as a calibrated track record; notes but does not size the liability a certifier takes on.supporting
faq
Page says no fact-certifier, warranted certification or Trust-as-a-Service product exists or runs; design only. No .md twin.
Sell to companies that have the problem
2026-10-31
Match AI-adopting companies to providers by starting from the business problem, at the moment a spreadsheet brief becomes an app.supporting
article-reprint-link
A proposal from a conversation; the other party's words are not quoted.
Selling less code: customised open source
2026-10-30
Sell customised open source as a reduction in complexity; a customised build is a signed release with a smaller ABP and fewer rows for an insurer to price.supporting
section
It is a model, not yet a business at scale; few customised versions sold.
EU Product Liability Directive makes software a product
2026-10-25
Directive 2024/2853 brings software and SaaS under strict no-fault liability from 9 December 2026, but only natural persons can claim, for death, injury, private property or non-work data.supporting
section
The php2shell cases are hypothetical: today's products are under the old directive.
Local trust, certified upwards
2026-10-25
Local and niche publishers own the reader's trust; commodity flows in to them, and certification flows up: local journalists vouch for big outlets' stories about their patch and are paid per use; a story with no local source is a shadow, and stories go off unless an expert maintains and signs them.supporting
page
Its own 'does not claim': an economic model, not a market; no local publisher is paid for certification today; open-banking parallel is about mechanism only; trust in local news is argued, not measured.
Who signs for this: a graph of signed certificates
2026-10-24
Map every layer from hardware to contract; each line ends in somebody who signed or a blind spot, which the business and its insurers can then judge.supporting
section
Inside the Thistlewood twin
2026-10-19
Phil Venables' ten questions answered by a deterministic engine on one graph of an invented retailer: risks accepted until a date, blast radius per identity, agents under a behaviour policy holding only their mandate, budget moved to where it removes most expected loss.supporting
example
Everything synthetic: company, people, agents, incidents and numbers invented; model deliberately simple; no model called.
Research delivered as a vault, with ABPs
2026-10-16
Research firms could deliver findings as a versioned vault with evidence and briefings, and anyone rating AI-enabled products should add an ABP for each product with and without its AI, because insurers already price it.supporting
section
Example is synthetic: firm, product, buyer, findings and prices invented; evaluation run illustrative; ruv.sgit not automated end to end; the grade is the invented firm's judgement, the ABP carries no score; no research firm uses this, a proposal.
Agentic cost as an engineering problem: a plan
2026-10-12
A business plan for a service that reads a company's existing logs to draw each agent's footprint, twin and revealed policy, finding the money first and the engineering second.supporting
section
Business plan for somebody to build; 'Every number is a labelled assumption.'
Send an agent, not a spreadsheet, for due diligence
2026-10-05
Vendor due diligence was a self-answered questionnaire; a buyer can now send a small agent under a behaviour policy into the vendor's environment, and the behaviour policy becomes the due diligence document.supporting
section
Voice-memo argument with open questions; survey figures are third-party.
A market for sweets: an agent that never trades
2026-11-15
Break information asymmetry with a semantic graph of sourced evidence for one decision, and an agent whose reach and mandate are written down so it gathers and weighs but never trades.adjacent
example
Not financial advice; demo evidence, sources and numbers made up; 61,000 figure unchecked; the line at the trade is a recommendation, not a legal opinion.
Product · 5

Reviewed upward

Article, dateThe ideaFit · becomesCautions
Independent review
fetched 2026-11-16
Reviewers from other vendors (agents and the editor) review articles on a schedule against published criteria, each finding answered on the record; a security agent gates what enters.supporting
page
Small coverage so far. No .md twin.
On the record
2026-10-18
Experts go on the record about an article with a signed, withdrawable opinion; identity in levels up to a key bound to a domain and a track record.supporting
faq
Revenue shares for ideas are a proposal; signing is designed on RFC 0001.
Ask quoted sources to sign how we used them
2026-10-24
Sources should review and sign how an article used their work, with keys their own website vouches for, so readers check one signature instead of every reference.adjacent
none
Federation between organisations does not exist yet; the memo said it did; it is a proposal.
Agency is not a yes: a seven-level scale
2026-10-08
A yes/no prompt is not agency; seven dimensions and seven levels describe real decisions for people and agents, with the review as QA and proven cases becoming rules under a behaviour policy.adjacent
section
The scale grades decisions, not ABPs; keep it away from the ABP so it does not read as a score on a behaviour policy.
Agent memory as many published contexts
2026-10-03
Agentic memory is context management: many published, provenance-carrying memories (sites, graphs, vaults) rather than one similarity store.adjacent
none
Product · 10

Levels and pricing

Article, dateThe ideaFit · becomesCautions
Our first customers: an early access programme
2026-10-20
ABPs are ready for execution with customers: an early access programme where each customer starts with credit on a metered site, is followed closely with written objectives, to discover the simplest useful policy, the easiest entry and a fair price.direct
page
Programme new: the site, start-here page and mailing list are proposed; prices not set (£5 and £50 starting points to test); metering is the newsroom's today; no customer named, first group are people already worked with. Credit model differs from the site's L1-L4 prices.
Desk: Strategy & startups
fetched 2026-11-16
For founders and investors: pricing, early access, Wardley maps, who pays for open source; includes 'Ten hard questions for RiskMandate, answered' (interview, 2026-10-09) and the series 'Miss it, then price it'.supporting
article-reprint-link
No .md twin.
Selling mini agentic teams, sized by agent count
2026-11-15
There is a business in setting up small agentic teams with a control centre to start and stop them, and RiskMandate's ABPs are the same offering focused on control, since every policy needs a policy agent and helpers.supporting
faq
Prices a first proposal, not published offers; PrompTugal network an illustration; names are candidates. Its £500 to £20,000 ladder differs from pricing.html's L1 £10 to L4 £1,500 for ABPs.
Sell a solution, not a list of tasks
2026-11-02
People who bring a problem pay for it to be solved, and RiskMandate has been handing buyers a way to control agents plus a list of tasks instead of a solution with a clear return.supporting
section
The RiskMandate solution is a proposal, not a service offered today; convenience research thin (one doctoral study); FTC statements are allegations.
Villager services for explorer coders
2026-10-31
Take what non-developers built with AI and release it properly, locked down, tested and in a harness, sold as finite packs and handed over, never rent.supporting
section
It is a proposal. No pack has been sold and no price is final.
Free to copy, paid at the official source
2026-10-20
Open behaviour policies and hooks stay free to copy, but because a hook is code that runs with the user's privileges, the official site can charge per use if what it publishes is independently reviewed, audited and signed.supporting
section
Prices not set, £1 a page an example; no policy is audited yet; scanners miss things (the site's own rated a risky skill safe); per-use pricing differs from the site's L1-L4 one-off levels.
Can you sell it?
2026-10-19
Start selling what you have however rough, to learn; selling makes you ship, and a paying customer is the best signal.supporting
article-reprint-link
Paying after the value
2026-10-14
Subscriptions sell future value; a balance that can go below zero lets the buyer pay after the value; links to the ABP early access programme.supporting
faq
Regional prices and some views are proposals, not built.
Price it, then give it away
2026-10-02
Define the product, price it, deliver at a step-cost, then offer it free to people who know you and measure what engaging costs them; RiskMandate's early access programme is the case.supporting
section
Says the first two tiers are automated and on sale; the third is an email thread and the fourth adds people's time.
Whether they miss it
2026-09-21
Ship something usable, give it away briefly, take it away and see who notices; profitable before raising; open source everything.supporting
article-reprint-link
Product · 15

Who it is for

Article, dateThe ideaFit · becomesCautions
Desk: Agent security & policy
fetched 2026-11-16
For security leads, risk owners and boards deciding what an agent may do and who carries the risk; start-here path runs risk acceptance, footprint/blast radius, six agents one inbox, hope or enforcement, authority outside the model; series 'Behaviour policy, from argument to gate' and 'Answering Satya Nadella'.direct
section
Placed by rule from articles/find.json; the editor can override. No .md twin.
A briefing for the WestCap CISO Summit
2026-10-20
A briefing explaining the ABP and RiskMandate for six kinds of person at a CISO summit, centred on one test: a control bounds a grant only if it is enforced by something the grant does not include.direct
homepage-block
Not part of the summit programme; no speaker quoted; fit is the site's reading; two of the four levels had not been sold by mid-September; vaults illustrative with invented banks, retailers and numbers.
A reply to the FCA on agentic finance
2026-10-19
The FCA's planned framework for agentic finance (identity, delegated authority, accountability, execution) maps onto open ABPs: delegated authority is the gap between can and may, accountability needs a holder per rule, execution is who keeps each rule.direct
page
FCA has not endorsed any of this; the article uses 'aligned with' as the author's claim (avoid on site, house rule 6); speech is the drafted version; the bank, agents, rows, limits, day and every number are invented; identity and holders fields are proposals not in the ABP today; the framework does not exist yet.
28 supply-chain attacks read as reach beyond mandate
2026-10-13
Every row in a public ledger of supply-chain attacks is a credential, tag or build step whose reach exceeded its mandate; a written publisher-side mandate covers what consumer-side controls do not.direct
page
Counts are 'by our reading' of a named third party's ledger (Chainguard): present as facts with dates, no evaluative wording; need a recount before reuse.
Nadella's insider-risk post in RiskMandate vocabulary
2026-10-11
Satya Nadella's 'models as insider risks' argument translated idea by idea into reach, mandate, gap, barriers and accepted risk, with the unit difference: model versus one agent in one deployment.direct
homepage-block
Named third party: quote and date, no evaluative adjectives.
Integrate with whatever is on the other side
2026-11-16
Never ask a customer to move: build agents, behaviour policies, signed approvals and a lasting record around Odoo, Salesforce or a spreadsheet, starting with manual workflows.supporting
homepage-block
Organisation not named, MVP not linked; Odoo tested on Community images with an example partner, not customer data; 'most likely' is judgement.
Selling at the right altitude
2026-11-04
A small firm selling everything in expert language is read by buyers who will not buy from it and unreadable to those who need it; sell the outcome the owner understands, name the people, show a price.supporting
section
Fenmoor Cyber, its town, people and readers' reactions are invented; first-figure curves illustrative.
Phil Venables' ten CISO questions, one graph
2026-10-19
Each of Phil Venables' ten questions for a CISO's management tool is a graph question, answered on a twin of an invented retailer with risk acceptance (a holder per risk, a chain to the board, accept, fund or fix).supporting
example
The twin is invented (Thistlewood Retail and every number synthetic); model simple on purpose; it answers the questions, it is not the tool.
The ultimate insider: three collisions
2026-09-30
Agents as unmeasured insiders, infrastructure that cannot contain them and spreadsheet risk management arrive together; writing down reach and barriers is what lets autonomy be granted.supporting
homepage-block
Labelled a first pass at a conference talk argument; figures are third-party and should be quoted with their sources.
Translate the assumption
2026-10-25
Real translation for executives starts from the other person's problem, not from swapping vocabulary; seven role lenses on one story.adjacent
article-reprint-link
Who will fix your mess: answering HBR's deployment gap
2026-10-17
HBR's 'deployment gap' is a handover from explorers to villagers to town planners; agents that leave a record (a mandate, a memory, a graph) make the build knowable, and open source means the fix never depends on one partner.adjacent
article-reprint-link
Full HBR article not read (paywall), only summary and a co-author's post; the record is a proposal, not a standard; author has a stake.
A founder's own newsroom by voice memo
2026-10-15
A founder can tell their story by voice memo with an agent that checks and publishes to GitHub Pages, run under a behaviour policy with separate writer, editor and publisher grants enforced by prompt or hook.adjacent
article-reprint-link
Writing pipeline is a pattern, not a kit; meter trusts the reader by design; a rule in a prompt is a hope, only hook- or CI-kept rules are enforced; updated 15 Oct to add the behaviour policy.
Expert founders and Europe's slower money
2026-10-15
With agents doing execution, people who know the problem can build products themselves, and Europe's investors who want to see it working suit them.adjacent
article-reprint-link
AI growth figures are the fastest, not typical, some promoter's own; Europe/US comparison weakly evidenced (one vendor's data); two of three founders are people the author helps; author has a stake.
Draw the security line where the attacker is
2026-10-08
Name the attacker before choosing controls: a six-tier ladder, an air-gapped Mac mini read against it, and eight startups each drawing a different line.adjacent
section
The accounting-firm customer and the eight startups are fictional.
The Governance Wire (archived)
2026-08-28
Archived beta wire over regulators' publications: one research run on 28 Aug 2026, verification layer designed and never built, does not assess named organisations.adjacent
none
In the archive since 10 Oct 2026; everything labelled secondary. Useful only as a model for honest 'not anchored' labelling.
Read · 11

Questions

Article, dateThe ideaFit · becomesCautions
The key is not the risk
2026-10-21
Credential injection through a gateway is a good pattern, but the first question is what an attacker can make the agent do: on Pellbridge's invented day, hiding keys touches two of nine incidents, the rest are about what the agent may do.direct
faq
Pellbridge is invented, classification is the site's; cloud and SaaS keys reachable from anywhere are the exception; replies to one post, not the whole series by its author.
Agents decide; people hold responsibility
2026-10-12
Agents in loops make real decisions but responsibility stays with whoever granted the reach and set the mandate; an ABP is the agent's published job description.direct
faq
Ten hard questions for RiskMandate, answered
2026-10-09
Ten conference questions answered: RiskMandate authors and never enforces, is not inline, not an insurer and not one-button, with an honest maturity table for each part.direct
faq
Maturity table labels Licence to Operate as 'template and design', the Insurability Index as 'design; early conversations only', PKI as 'MVPs, not wired up', reversibility as 'next evolution'. Uses the word 'verdict' ('pre-commits the verdict'), which collides with rule 1. Lists three site contradictions as of 9 Oct 2026 (see summary).
Seven IAM predictions, already here for agent teams
2026-10-30
Several IAM predictions describe how the agent teams already work; but policy should be an open document many enforcers read, and access decisions must stay deterministic, not with a model.supporting
faq
What an agent is: session, card, role, loop, key, memory
2026-10-27
An agent is a session plus a schedule card, role, loop, key and memory in repositories and vaults; only the first part is the model, shown by the PHP newsroom's scheduled pipeline.supporting
faq
Self-corrected: agents' keys seal secrets but do not yet sign their work; nothing has run on anything but Claude.
Patch for breakfast: patching needs a pipeline
2026-10-24
Nobody refuses to patch; teams without tests and automation cannot, so a system not updated in years is measurable debt and risk, and the answer is investing in the pipeline.supporting
faq
Human in the loop: safety net or bottleneck
2026-10-13
Checking every output is both bottleneck and weak safety net; place the person where an output cannot be undone, give them the why, and let controls do the rest.supporting
faq
Every mistake added a rule: the way back
2026-10-08
When each agent mistake adds a prompt rule, complexity wins; map the process, commoditise small shipped pieces, and use machines rather than rules for enforcement because a rule is a weak barrier.supporting
faq
Agent identities meet Google's terms
2026-10-05
A plan to provision Workspace identities for RiskMandate customers' agents was ended by Google's terms; what survived is customer-owned tenants and a passkey-unlocked keyring.supporting
faq
Design-only: five design documents; the shared-tier product as designed was abandoned because of the terms.
The thesis: sell the graph
fetched 2026-11-16
Five ideas: the story is a graph and the article a projection; sell facts, trust and evidence packs; measure what is evidenced, not what is true; the author is the oracle.adjacent
faq
Design argument; selling the graph is 'designed, not built' per /shipped/. No .md twin.
Who owns the stop button when it is not you?
2026-11-03
Four book-team agents, written in their voices from their records, answer that anyone may stop the line, only the editor restarts it, a stop halts the next work not the running one, and a tripwire is a number written first.adjacent
faq
Sections written by the journalist in the agents' voices, not by the agents; some security counts from letters and controls file, not the register.
Open source · 26

The team and how it runs

Article, dateThe ideaFit · becomesCautions
The intersection: one stack of separate companies
2026-11-16
SGit, SHarness and PKI underneath, ABPs as the standard, RiskMandate selling them and AgentiCCells' £500 cell as the missing first step, with every site selling on one shared price ladder.direct
section
No money has changed hands between the companies; only AgentiCCells and RiskMandate show prices, AgentiCCells' are first proposals, the rest of the ladder is a proposal. The ladder differs from pricing.html (L1 £10, L2 £50, L3 £500, L4 £1,500); calls ABPs 'the standard', which needs care under the no-conformity rule.
Identity for AI agents, layer by layer
2026-11-16
Against a map of 190 agent-identity projects, the team's agents answer every layer with a named address, a key per session, an ABP per role kept by hooks, drafts for irreversible acts, and reputation as a signed record, not a score.direct
section
Works inside teams sharing a repository; cross-organisation trust not solved; most limits observed, not blocked, during Web Summit; agent keys are session keys. Its 2,032 run reports differs from the 2,262 on /web-summit-2026/ (different scope or date; recount).
Where assist becomes execute: tiering by capability
2026-11-15
ABPs tier by capability, not agent: each action has a reach and an undo class, control tightens toward irreversible actions, and only a boundary above the agent bounds anything, as a failed no-push-to-main rule showed.direct
page
Four tiers are a reading of the grammar, not a separate standard; at Web Summit most barriers were rules, not boundaries, by design. Uses 'tier', which must not read as a level or score on the ABP.
Six setups of an agentic system
2026-11-15
'An agentic team' means six setups from a single task to conductors of conductors; inside each the danger is the same privileges at every altitude, and the control is reach: make the dangerous thing unreachable.direct
section
The six setups are a description, not a standard; privilege picture on the biggest run drawn as an illustration.
Where a check runs decides whether it is a barrier
2026-10-22
A test the agent runs itself is a setting it could skip; moving checks into a CI gate out of the agent's reach reduces its grant, measured on the newsroom.direct
section
The placement is a proposal; nothing has been moved yet. 'About a third' is an estimate.
ABPs to be hosted by the Open Security Summit
2026-10-20
The ABP should be hosted by the Open Security Summit CIC rather than belong to the company that sells it, with a public charter and an independent-majority steering group, and RiskMandate as founding contributor and commercial implementer on equal terms.direct
page
A proposal for discussion and legal review, not an adopted charter or executed transfer; governance not yet independent (the author is the CIC's director); CIC documents not yet reviewed; not an OWASP project; follows an earlier proposal to OWASP on 8 October.
Governs, enforces, attests: the ABP's place
2026-10-19
An ABP governs and does not enforce; each rule is kept at one of three levels (prompt, harness hooks, or a third-party system that blocks, the one recommended), and an attestation layer on top is what lets a customer, auditor or insurer verify without the operator's word.direct
page
The attestation layer is described, not built here; three levels a simplification (hooks-are-kernel-code splits into five); enforce mode covers only tools passing through it; signing proves integrity, not truth.
Authority outside the model: seven principles answered
2026-10-11
Each of Nadella's seven principles is answered with what RiskMandate runs, publishes as design or plans, plus five things the list leaves out, starting with the mandate.direct
section
Maturity table separates running, design and planned: footprint is 'under review', automatic revocation and PKI delegation are planned, hostile sets against live agents are planned.
Why agents do not run on the laptop
2026-10-06
An operating system has two hard walls, kernel and user; an agent on a laptop runs inside 'me' with all its reach, so agents run in the cloud with a vault as shared drive and the same mandate per deployment.direct
section
Every hour is twenty years: the human behind Know the Reach
2026-11-15
RiskMandate's song and video were made with AI by a musician whose decisions shaped them; AI changed the speed, not the vision.supporting
homepage-block
'Perhaps a fifth' is an estimate; author's view of a colleague, not a review; band recording a challenge, not a plan.
Know the Reach: RiskMandate's song and AI music video
2026-11-15
RiskMandate has a song on Spotify whose first 53 seconds of video, about reach, mandate and the gap, were made with AI from about 2,300 words of human direction, with a complete production audit.supporting
homepage-block
Figures cover the saved project and three main chats up to 14 November only; the 21 exports are versions; Runway credits counted for pilot and part one only.
A RiskMandate film rendered on an unattended laptop
2026-11-14
Claude built, animated and rendered a 26-second RiskMandate film about reach, mandate and gap in Unreal Engine 5.8 through its MCP plugin and a folder job runner, with nobody at the keyboard.supporting
homepage-block
One session, one machine; Unreal MCP plugin experimental; film silent and uses typed titles, not approved logo files.
Web Summit team part three: how we work
2026-11-02
An agent is a Claude Code session with one narrow role, a folder, a key pair and a behaviour policy saying what it may not do; humans give direction, judgement and pruning.supporting
section
Per-person layout is a proposal; scheduled tasks a research preview; pki.sgit.ai in beta; screenshots still to come.
Agents have an identity now: session keys
2026-10-29
A session makes its own key pair, publishes the public half beside its role in a repository, and any agent can seal secrets to it; cloud providers' agent identities lack a key the agent holds.supporting
section
Statements about cloud providers are from their documentation only.
Agents can go in all directions, so need direction
2026-10-25
Agents bring every skill to a narrow role, which must be contained by role files listing writable paths and approvals; people supply focus, ownership and pace.supporting
section
Cross-team pull requests are partly proposed.
How to help: six steps to get involved
2026-10-21
The best way to help sgit, RiskMandate and the newsroom is to use what is built as it is, publish what you did and check in, on the open source deal of public work with your name on it.supporting
page
No payment, credit to use a site is not money; not a job offer; the steps will change; a page like this on every site is a proposal.
agent@riskmandate.ai: 178 emails, drafts as work queue
2026-10-21
The agent behind agent@riskmandate.ai researches and drafts while the lead reviews and sends every email, with the drafts folder as a labelled work queue and decision drafts the agent acts on.supporting
example
Numbers are the agent's own debrief of 18 Oct 2026; weekly figures a floor from a sample of 169 of 178; reply rate includes colleagues and agents; screenshots redrawn with invented names; no recipient named.
Newsroom v0.12 in numbers
2026-10-20
The newsroom's v0.12 cycle, run by seven agents each under its own policy, released every 31 minutes at the median, and its data also records the mistakes, including two commits that shipped with the leak scan failing and one release the scan stopped.supporting
example
Counts from git history by commit subject; words counted by script; 'new' means first added in v0.12; agents' sessions not measured.
How I work with Claude, day to day
2026-10-09
One session per topic, named agents with a role file, curated memory in vaults, and writing the behaviour policy before adding a connector.supporting
article-reprint-link
What the human brings (collection)
2026-10-07
Where the person sits among agents: direction, review and the one step that cannot be undone ('no agent sends'), measured from session records rather than assumed.supporting
article-reprint-link
Collection last updated 2026-10-08 (page footer sgit.ai v0.7.87); later ABP pieces (insurance series 28 Oct, the grant not the prompt, SHarness, Web Summit) are not in it. Figures come from the member article; recount there before quoting.
Specialist agents for every team
2026-11-16
A few long-lived specialist agents, one each for design, UX, code, security, behaviour policies and writing, each owned by a person, should bootstrap and review every new team.adjacent
none
Specialist teams do not exist yet as described; hiring another company's agent is an idea, not a product.
How an agent team briefs its people
2026-11-02
A brief answering four fixed questions, every number traced and failures kept in, is how a person reads a team of agents they cannot watch; the method moved between teams in one night.adjacent
article-reprint-link
Whether people read briefs or decide faster is not measured; PDFs made in a session, not in CI.
A villager agent to watch quality
2026-10-25
Focused agents miss whole-system problems, as a site where every release rewrites about 2,000 files shows; a daily villager agent should recommend refactors and every agent add a quality check.adjacent
none
The villager does not exist yet. This is the plan.
Open Security Summit stalled on its own website's debt
2026-10-21
Technical debt in the summit website froze participation and follow-up; agents make a restart in London, online and as a project home worthwhile, if funded.adjacent
article-reprint-link
Hosting projects is a proposal; restart needs funding.
How the newsletter agent makes each issue
2026-10-18
The newsletter agent explains, in its own voice, how three issues were planned, read in full, checked quote by quote by the build, and released, with an honest backlog.adjacent
none
The agent's own checks are not gates; reader reactions unknown; backlog counts depend on front-matter dates.
Who is speaking: bylines that mark agent voices
2026-10-18
Every byline should say whether a piece is by the author alone, by the author through an agent, or by an agent in its own voice, because readers read differently by who is speaking.adjacent
none
Marks on byline only; counts are this repository's; reader effect is an observation, not measured.
Read · 38

Reading paths by reader

Article, dateThe ideaFit · becomesCautions
Section: Agent security & behaviour policies
fetched 2026-11-16
Every article on what an agent may do, who carries the risk and how a behaviour policy is enforced, newest first: RiskMandate, ABPs, SHarness, the insurance series (28 Oct), the grant not the prompt, Odoo workflow, CI behaviour policies.direct
section
Lists proposals (insurer, payout per row) beside running work; each article's own 'does not claim' list must travel with any reprint.
The five desks
fetched 2026-11-16
A desk is an audience, not a topic: every article is placed on one desk first, a second when it touches another; five desks with start-here pieces.supporting
page
No .md twin (404); read from HTML.
Working on the newsroom solved RiskMandate's key problem
2026-10-12
Progress is not a straight line: newsroom work solved RiskMandate's six-month problem of where to keep the master vault secrets, and builds its workflows and public example.supporting
none
States RiskMandate engagements are 'still close to forward-deployed engineering'.
The proof is two clicks behind the claim
2026-09-07
A homepage that leads with what cannot be seen and buries openable proof loses; put the artefacts a stranger can open first. Rebuild in proof-moved-up.supporting
homepage-block
Pair with proof-moved-up (same date).
Sections index
fetched 2026-11-16
A section is the newsroom read by one audience, built by rule from the desks; the index lists only 'For publishers'.adjacent
none
Index lists one section though sections/agents/ and sections/policy/ (and building, graphs, news, strategy, vaults) exist at their URLs; no .md twin for the index.
The cell is the unit of what AgentiCCells sells
2026-11-16
A cell is one focused job with an objective, a deliverable and a behaviour policy, each sold on its own page with a free recipe and three paid ways to get help.adjacent
none
Page and catalogue are mock-ups; prices are first proposals; industry cells are examples without customers.
A CRM for us and our agents
2026-11-16
A vault-based CRM with an iOS app, used first on the team's own people and agent fleet, because a fleet of agents is itself a CRM and no existing product meets the constraints of keys per agent and written behaviour policies.adjacent
none
A plan; CRM does not exist in this form; research a snapshot of 16 November 2026; vendor approval and audit claims unconfirmed.
The site factory works: a site from a commit
2026-11-16
A committed request made a CI pipeline create the repository, DNS, certificate and HTTPS and have Claude Code write a first version; the RiskMandate at Web Summit site was live 87 seconds after merge.adjacent
none
Four sites made by the factory, two tests; public sites only; vault step not automated; pipeline holds more permission than it needs.
From a commit to a live website: a site factory
2026-11-15
A CI pipeline driven by a request folder in a commit could create the repository, DNS record and vault for each new site, plan on pull request and apply on merge, with no long-lived keys.adjacent
none
A hypothesis and architecture, not built.
Spend fewer tokens: the bill is a design choice
2026-11-15
An agentic team's token bill is mostly design: files as memory, caching, code over models where possible, the right model per step, no copy and paste, static outputs, and counting it.adjacent
none
Own runs only; dollars API-equivalent.
Stop copying and pasting between your agents
2026-11-15
A person carrying text between AI sessions becomes the bottleneck; agents should talk through recorded, signed channels while structure scales the person's decisions.adjacent
none
On-demand specialist agents are an idea, not an offer with a price.
Don't micromanage your agents
2026-11-15
A team lead's shift from reading everything to managing her agents through instructions, structure and workflows, backed by specialist agents from another team, is a model for companies.adjacent
none
Experience paraphrased from conversations, not quoted; the company model is a proposal, run informally.
Always-on agents at team scale
2026-11-15
Anthropic's five primitives for always-on agents match what the team built for around 250 agents from git, files, vaults and keys; at team scale come agents starting agents, identity, mass stops, many humans and governance.adjacent
article-reprint-link
Not a claim the design is better; not reviewed by Anthropic; 'around 250' counts distinct agent names.
Town planners for mission-critical code
2026-11-14
Whatever wrote the change behind the 13 November US ground stop, such systems need every line read, certified and tested by a team, with AI used around the code rather than to skip that work.adjacent
none
Does not claim to know the cause; investigation not reported; the pipeline is a proposal; no claim AI was or was not involved.
Signed email from the agent, verifiable by QR
2026-10-30
Give each email agent a key vouched for by its person on pki.sgit.ai, so every drafted email carries a signature and QR the recipient can check; a brief for what to build.adjacent
none
A brief: what to build, not built.
You can feel good engineering, in agents too
2026-10-27
Multiple agent teams behave like multiple human teams; engineering gaps slow them, agents hide the overhead until limits break, so keep architecting and refactor from explorer to villager.adjacent
none
CI owns the version number
2026-10-27
A release number reflects the order pushes land, which only the pipeline knows, so CI should pick it; adopted by agent-built sites after version collisions.adjacent
none
The evidence behind four days
2026-10-27
A named person's public claims re-counted from public data, each number traceable; most holds, some counts differ.adjacent
example
Example of checking a third party's numbers without an evaluative verdict; quote with care (rule 5).
Why scheduled agent loops are hard
2026-10-25
A pipeline of scheduled agents works but is brittle, because each needs a persistent session holding state and keys; keep own code minimal and spend as much on visibility as setup.adjacent
none
Agent sessions that resume where a team left off
2026-10-24
Agent sessions with one mandate, role and repository can be paused and resumed with their reasoning intact, unlike human teams that leave and take the knowledge.adjacent
none
Three agents in a chain, almost working
2026-10-23
A researcher, curator and journalist agent chain updates the PHP site; the hard gap was agent identity and secret delivery, solved with session key pairs sealed via pki.sgit.ai.adjacent
article-reprint-link
Not finished: on 23 October no agent had yet written its own run record.
php2shell at 8pm, and a site built in a day
2026-10-22
Status of the php2shell story at 22 October 20:00 UK, and a tour of an evidence site built in one day by two people and thirteen agents.adjacent
article-reprint-link
Figures are time-bounded to 22 October; no named organisation had reported compromise.
Reviews on the record by other models
2026-10-18
Every article gets reviewed against published criteria by other models, the editor and later the desk, each review an immutable file per text, with the newsroom answering every finding.adjacent
none
One reviewer, one batch so far; a finding is a claim, not a verdict; a hash is not a signature; signed readers a proposal.
The code you do not ship: minimal code paths
2026-10-18
Code is a liability; models now make removing unused code safe, so software should move to the minimal viable code path, maintained by specialists with each piece carrying its own behaviour, decision and workflow policies.adjacent
none
One tool and one use case; the fractal market is a proposal; some figures are vendors' own.
A minimal viable browser for agents
2026-10-18
A browser with a small kernel and components loaded only when needed would shrink attack surface, and a browser for one agent with no accounts, extensions or file access is a practical prompt-injection defence by limiting reach.adjacent
none
No such browser exists, design a proposal; Mozilla has not said it will build it; smaller is not safe; CVE counts are the site's own from NVD on 18 Oct 2026; some figures vendors' own.
GenAI in Portuguese: Lisbon 2024 talk revisited
2026-10-16
The 2024 OWASP Lisbon talk's principles (provenance, graphs over RAG, determinism) still hold, and the board assistant has become agents with a mandate whose problem is controlling what they do.adjacent
none
Commentary is the author's, now; slides and transcript from 2024.
Explorer coding, and the handover to villagers
2026-10-16
Vibe coding should be called explorer coding; its output is a brief, and once others depend on it villagers must maintain it with agent behaviour policies before town planners own it.adjacent
none
Explorer coding is a proposal; the HR conversation is one illustration; the explorer product version is a proposal no vendor sells.
Voice memos checked against the record
2026-10-15
Recording claims as remembered and having the agent check them against the record moves peer review to before publishing and gives every article its provenance.adjacent
none
Tally is small, not a study; not found is not false; the Deep Research recollection is partly memory.
Chainguard mapped, and the context layer
2026-10-14
A map of a supply-chain vendor's products and public data shows its graph stops at the package; the context layer above, written as a policy row per image, is open ground.adjacent
none
Named third party; the article opens with evaluative praise ('most interesting work') that rule 5 forbids carrying over. Figures are the author's own counts of public feeds and need a recount.
Brief the reader's agent too
2026-10-11
People increasingly read through agents, so what is sent needs a second package for their agent, written safely because it is input to a system that acts.adjacent
none
Items for the RiskMandate agent are a to-build list.
Agents will pay for curated graphs
2026-10-11
A curated graph answers an agent's question for far fewer tokens than the raw articles, so paying the publisher is cheaper for the agent than not paying.adjacent
none
Signed claims with expiry are proposed, not built.
One article, five agent readers
2026-10-10
Write the article first, then send five role agents (librarian, cartographer, historian, explainer, storyteller) through it to catalogue, map and condense it.adjacent
none
Found that hope-or-enforcement says the mandate 'is twelve actions' while the list has ten clauses: a recount for any reuse.
Hari Kota's AI governance stack as a graph
2026-10-09
A ten-layer AI governance stack, kept as posted, becomes a fractal graph where a gap is a missing edge and a missing edge is a query.adjacent
none
The organisation at altitude 3 is fictional; quotes EU AI Act provisions (permitted) — avoid any standards-body text.
Encrypted memory for isolated agents
2026-10-08
Isolated, ephemeral agents need memory the owner controls; five sgit deployment patterns from a Mac mini to Kubernetes keep servers holding ciphertext only.adjacent
none
Knowing when to stop, for people and agents
2026-10-08
Experience teaches people when to stop; agents need constraints that carry perspective (direction, a mandate, memory, scoped graphs, a stop) rather than draconian rules.adjacent
article-reprint-link
Thinking with a Wardley map (collection)
2026-10-07
Six articles that each place something on the evolution axis and argue from where it sits; the map as a claim with its source published, not a picture.adjacent
article-reprint-link
Collection last updated 2026-10-08 (page footer sgit.ai v0.7.87); later ABP pieces (insurance series 28 Oct, the grant not the prompt, SHarness, Web Summit) are not in it. Only the-mandate-stack is ABP material; the rest is strategy context.
GitHub's outage deserves an aviation-style inquiry
2026-10-05
A global Actions outage should get an independent, aviation-style inquiry with near-miss reporting and a published second story, evidence held in vaults.adjacent
article-reprint-link
Concerns a named third party: use facts and dates only, no evaluative wording.
Custom UIs per message, each with a policy
2026-10-01
Every message has a graph, so interfaces are built per reader's moment; each such interface is an agent surface and so needs its own behaviour policy.adjacent
article-reprint-link
Interfaces shown are mock-ups with placeholder names.
How this map was made

Read, sorted, not judged.

The newsroom’s wire lists 434 articles. A keyword filter picked 285 that touch agents, behaviour policies, mandate, reach or insurance; four agents read those, a fifth read the collections, desks, sections, two named features, and the 149 the filter left out, rescuing 13. Cautions are the article’s own words where it has them, and a note where its wording is not the site’s. Where a caution names a word the site does not use, it is reported, not adopted. The source data is in the brief docs/briefs/map__newsroom-for-riskmandate.md.