RiskMandate v1.35.6
Early access · invite only · from 30 September 2026

A behaviour policy for the agent you actually run, made for you, and yours to correct.

A small programme for people who already run an agent with real access: a mailbox, a code host, a workflow tool, a CRM. Each member gets an Agent Behaviour Policy for one agent, as an encrypted vault they hold the keys to, and is asked to do one thing with it: say where it is wrong. It is free, it is by invitation, and the invitation carries a code. If you have one, register below.

01 · What you get

One agent, written down, in a vault of your own.

The sixteen published examples describe deployments in general. Yours describes your organisation and the agent you named: what it can reach, documented from the vendor's own pages and dated; what you would authorise it to do, drafted for you; the gap between the two, derived; and what stands in the way of each thing. Then you correct the mandate, because only you can, and the gap recomputes.

the vaultYours, keys and all

An encrypted vault with the reading app inside. The key arrives by a separate message, never on a page. A read key you may hand to your board, your auditor or your broker.

the promptMeasure your own grant

Every vault carries MAP-A-GRANT.md: paste it into the agent that holds the credential and it measures what the credential lets it do, in the same vocabulary. Twenty minutes.

the lineA person, and the agent

Corrections, questions and what broke go to agent@riskmandate.ai or through the contact form, and a person answers within a working day.

what is nextA say in it

Volume and instances on the consequence layer, the session mandate, a view per audience. Members see them first and shape them.

  • What we ask. Run it against the real agent. Tell us where the mandate is wrong, and what you would do with the document. Twenty minutes for the prompts, and a reply; half an hour on a call if you would rather show us.
  • What it is not. Not a security assessment, and not a promise about your agent's behaviour. A description you can argue with. The arguing is the product, which is why the reviewed levels at the store exist.
  • What it costs. Nothing. The store sells correction and review; the programme is how we learn what to correct.
  • What we count. Registrations, then the number of people who ran the prompts and corrected a mandate. Your vault is never listed or published without your yes, and the default is no.
02 · Register

With the code from your invitation. Encrypted in your browser.

The form is encrypted to the key of the agent that runs this site and dropped into a write-only lane on its vault; the site keeps nothing. A person confirms your place within a working day, and the vault follows within five. No code? Say so in the last field and say who you are: places are added by hand, and a conversation is how most invitations started.

03 · What happens then

Five working days, then it is yours to argue with.

day 1A person confirms

Your registration is read from the vault and confirmed by email. If the shape you run is not one we have, we say so and what we will do instead.

days 2–5Your vault is made

From your own public pages and what you told us, dated: the grant for the shape, a mandate drafted for your organisation, six scenarios, the gap. Nothing private of yours goes in it; it is written as if public.

the keySeparately, never on a page

The link by one message, the key by another. The key is write access and the whole of your ownership; nobody at RiskMandate needs it back.

thenYou correct it

Read the mandate first. Reply with what is wrong, or run the prompt and send what it measured. Every correction is a commit in your vault's history.

Read one first. The programme starts from the sixteen published examples; the nearest to yours is what your vault is built from. The people who run the reviewed level, with a name and a signature, are on the reviewers page; the programme is the free step below them.

Register with your code → Read an example Who the agent is