Reach
Measured
Everything the agent can actually access in this deployment, including what nobody thought to check.
Agent Behaviour Policies for the AI agents in Gmail, HR, payroll, CRM and meeting apps: what each one can reach, what you authorised it to do, and the gap between the two. Enforce them with the controls you already run, or with SHarness, the open-source secure harness we maintain. We are never in the request path.
On Claude or ChatGPT, in your browser, nothing to install. Six examples, simplest first.
An agent, everything it can reach, the mandate the business gave it, and the gap between the two. Then the same record for every agent, read by operations, leadership, security and the insurer. No narration needed.
Film by Silvio Switha, November 2026, also on the Web Summit site. The film says delta where this site says gap: the same thing.
Blocked by your gateway: Payroll copilot tried to change bank details for 1 employee. Reply 1 to review.
The gate in the middle is your gateway, proxy or IAM, or SHarness on the machine the agent runs on, configured from the rules in the ABP. RiskMandate writes the rules and is never in the request path; the approval card is a flow your team stands up with our help.
Ask your assistant to do something. Paste one line that says it must not. Ask again, and watch it refuse. That line is the first row of an Agent Behaviour Policy.
In each example the assistant keeps the line because it chooses to: an expectation, the weakest of the four barriers. The examples show why the stronger ones matter.
One policy for each system your team already uses
Every system starts from a template. Decide what agents can do, what waits for approval and what never happens. We write it down; your controls enforce it. The rules you can paste into an agent today, one worry at a time, are here.
Upload the document legal already wrote. We turn every line into a rule, supply the business logic behind it, name the control in your stack that would enforce each one, and flag what nothing can enforce as written.
1Staff must not use AI tools to send confidential data outside the company.
2Customer data exports need a manager's approval.
3AI note takers may not record HR or legal meetings.
4Use AI responsibly and ethically at all times.
Instructions inside an agent can be ignored or talked around. A control outside it cannot, and the ABP says which of the two each rule has.
"Never export customer lists without approval."
Every rule in an ABP is kept by something. Told, it is a hope; checked, it is a hook on the agent's own machine; held, it is a boundary the agent cannot reach. SHarness is how a row gets held: an open-source secure harness, which we maintain, that you run in your own environment. We are still never in the request path.
The model keeps it because it chooses to. An expectation, and the one the five-minute examples show.
Deterministic, and on the agent's own machine, where the same account can switch it off. A setting.
The ABP becomes a Process Behaviour Policy, named by a hash you supply, and the kernel holds it on every process the agent starts. A tricked agent cannot widen it.
sharness-ai hello writes a PBP, hashes it, lets the kernel hold it, then runs six probes against it: six of six held, 25 October 2026, for a child process. The launcher that would hold the agent itself is still the design. Three more animations, from the gap to the radius, are on the SHarness page. E11, from the-ultimate-insider.presentation.sgit.ai, CC BY 4.0, copied here (credits).hello: six probes, six held, in five environments, and in a Linux kernel in a browser tab.An approval flow your team stands up with our help, from the rules in the ABP: the right person gets a text and an email, and approves or denies with a one-letter reply. It runs in your stack, not ours.
Held: Sales agent (for Dana R.) wants to export 2,400 contacts. Your CRM policy holds exports over 500. Reply A to approve once or D to deny.
D
Denied. Dana and the agent have been told why. Decision 48213 is in your log.
What is for sale is an Agent Behaviour Policy for the agent you actually run. Each level is the level below plus exactly one thing, and the one thing is what the price is for. Pick a level to see what arrives.
Prices are in pounds, per agent, per deployment, paid once at the store; no subscription. These four are what we can deliver today. Level 3 is where most people start, and it is the one we are doing with early users at no cost: register for early access, say which agent, and we build it and correct it with you in exchange for telling us where it is wrong. What comes next, and at what price, we will learn from those conversations. The full table is on the pricing page.
Not what they did. What they can.
An Agent Behaviour Policy writes down what one agent can really reach, what you authorised it to do and the gap between the two, in one record the CEO, CTO and CISO can all stand behind.
From £10 for one agent. All sixteen published examples are free to read.
The same model can be harmless in one setup and serious in another. Connect it to a mailbox and the narrowest Gmail scope that reads one message reads every message. So a behaviour policy describes one agent, in one deployment, in four parts, and carries no score. See the scope evidence
Measured
Everything the agent can actually access in this deployment, including what nobody thought to check.
Elicited
The job, written down: what the business authorised it to do. The one part only you can supply.
Derived
Reach minus mandate. Never written by hand, and recomputed whenever either side moves.
Recorded
What actually stands in the way of each capability: controls, constraints and open questions.
The files and the data keep the older names for these: GRANT.md is the reach and DELTA.md is the gap, because the vocabulary is pinned at abp.sgit.ai and sixteen published vaults are built against it. Same four objects, plainer words on this page.
The behaviour policy lives in an encrypted vault you hold the keys to, with a reading app inside. Hand a read key to your board, your auditor or your broker and they see exactly what you see.
Leadership view
“What have we authorised, and who owns it if it goes wrong?”
The gap in plain terms, a named owner, and a trigger that brings the decision back for review.
Operator view
“How do we keep shipping agents without losing track of what they can touch?”
Terms your agent reads in its own context, as AGENTS.md and SKILL.md. Markdown for people, JSON for tools, and nothing in your request path.
Security view
“Can we prove what it can reach, and what actually stops it?”
Every capability with its barrier, the gap recomputed whenever an input changes, and every version kept.
Insurance view
“What exactly are we being asked to cover?”
Explicit scope and the questions still open, so renewal starts from a record instead of a questionnaire.
It’s a draft on purpose. The people who built the agent, own what it touches and answer for it each correct the part they know. The correction is the product.
Three steps, in order. Each one needs the step before it.
Agent Behaviour Policy
What one agent can reach, what you authorised, the gap between them and what stands in the way.
See a behaviour policyLicence to Operate
The business is the authority, the behaviour policy is the instrument and the agent is the licensee, for a set interval.
See the templateInsurability Index
The record an underwriter will accept: scored, dated and with the residual risk owned.
See the designMeasured against the standards underwriters are adopting: ISO/IEC 42001, OWASP Agentic Top 10, NIST AI RMF and ISO/IEC 27001.
Model-drafted and marked as such. Not a compliance assessment. We are not an insurer and place no cover. What we don’t claim
Seven of us and our agents, each agent with its own published behaviour policy. Bring one agent you run, and leave knowing what it can reach.