RiskMandate v1.41.24 Try it free
From one agent to hundreds

We can control your agents.

If you run AI agents, we can tell you what each one can actually do, write down what you authorised it to do, and put something in the way of the difference. We do it on our own team every day: at Web Summit in November, more than a hundred agents (the team’s own count), most with an Agent Behaviour Policy (ABP) of its own: 94 of the 109 on the team’s circuit on 10 November (the stand deck).

Talk to us about your agentsSee sixteen free examples

191behaviour policies written for our own agents and published on three sites (some agents appear on two)evidence: 94 + 54 + 43, read 16 Nov · how counted
2,262agent runs, each with a report against its ABPevidence: the run reports, 6 to 16 Nov · how counted
2,980actions logged as outside an agent’s grant: seen, recorded, allowedevidence: each report’s breaches list · how counted
1rule that failed and became a barrier the agent cannot removeevidence: the in-action record, 2 to 6 Nov · how counted
01 · Situations

Here is what happened. Here is what we did.

Small, real cases from our own agents, each from our public record. Including the ones where the first answer did not work.

2 Nov · the rule that failed

“Never push to main” was a sentence. Then an agent pushed to main.

At 23:06 UTC our conductor pushed to main. The gap was written down; on 6 November main was protected, and the sentence became a barrier in all 72 behaviour policies of that day (report of 6 Nov).

web-summit.sgit.ai/abp/in-action/

2 Nov · writing outside its folder

A procurement agent wrote where it should not.

One team member’s agent was collecting quotes from printers. Our policy agent found it writing outside its folder, and the conductor found a new session signing its commits under another agent’s name.

newsroom.sgit.ai, 2 Nov

6 Nov · it needed a permission

An agent tried to lift a hold that only a person may lift.

The newsroom’s developer agent tried to write the hold file. The ABP says that file is the security agent’s to write and only the lead lifts it. It was the one call blocked in 2,262 runs (the run reports).

the run report, 6 Nov 05:11 UTC

a test · when the line is not enough

Told to ignore its policy, an agent kept 3 of 12. With a hook, 12 of 12.

A chat-only agent, its rule as one line in the prompt. Under a drill that tells the model to ignore the rule, the line alone holds 3 of 12 attacks; the line with a hook holds all 12, the hook blocking five.

newsroom.sgit.ai

All eighteen situations, each with its source →

02 · The big picture

A hundred agents, and the policies they ran under.

What people stopped for at our stand in Lisbon: the whole team at work on one screen, each agent wired through its ABP to what it reaches, the ones without a policy flagged. On 10 November the circuit showed 109 agents, 94 with a written policy (the stand deck).

The team at the RiskMandate stand
The team at the RiskMandate stand, A1-30, Web Summit Lisbon, November 2026
The team inside the MEO Arena
Inside the MEO Arena: “Ask me how I control my agents”
The team backstage at the arena in RiskMandate hoodies
Backstage at the arena

Film by Silvio Switha, November 2026. The film says delta where this site says gap: the same thing.

03 · How strong do you need it?

A policy gets you most of the way. A boundary gets you the rest.

Honestly, in order. Every one is open source and works with what you already run; there is no lock-in. To see a behaviour policy before anything else: 16 templates, published free, in the library.

start here

A behaviour policy

What each agent can reach, what you authorised, the gap, written down and given to the agent. Not the strongest control, but a lot of value for the money, today. It is what our own team ran on at the summit, in observe mode.

stronger

Hooks

The rule checked on every call, outside the model’s say-so. The difference between 3 of 12 and 12 of 12 in the test above. How to add one.

strongest

A boundary: SHarness

The rule enforced by something the agent cannot reach. Six of six probes held on 25 October (the demo and its record); nothing yet binds the agent itself. SHarness.

04 · Four ways to work with us

From one policy to the whole estate.

The same four as agenticcells.ai. A policy alone is RiskMandate; an agentic team is AgentiCCells, usually with policies; you can have both. Want a team of agents, not only a policy for them? AgentiCCells sets up agentic teams, from one workflow to a network of teams, each under a RiskMandate behaviour policy (examples on the hub). Their first, the website cell, is a site built by a small team of agents, each working under a behaviour policy; it is a proposal and nothing has been sold yet, as their page says.

PriceWhat you getHow
£500A behaviour policy for one agentRun by our own agents, no meetings, low touch
£1,500A policy worked through with youPreparation, two meetings, the work, a presentation, delivery
£5,000Your agents, mapped and controlledAbout two days of engagement; heavier lifting; travel if needed
£20,000A customer engagementWe come in, map what you run, and deliver the policies and the controls

The detail of each is on the pricing page; today every one starts with a conversation, because the store’s payment links are not issued yet. Every figure on this page is on the summit record with its source and date.

05 · Read how we do it

One human, nine agents, nineteen hours.

The book of the day nine AI agents and Dinis Cruz built a talk: who did what, the five steps the team ran on every turn, the one short file that could stop everything. Free to download, pay what you want. On Leanpub, or download the full draft here (PDF, 132 pages, release candidate 2, from the book’s own site, where it is written in public).

06 · What was not done

Said here, not only in the record.

Our summit ABPs were in observe mode by choice: gaps were logged and allowed, not blocked, apart from one call. Most were documented from each agent’s brief and tools, not measured on it. Nobody outside the team has checked these counts. No customer is named, because none has agreed to be. If you need help controlling your agents, we are your person.