Govern

Most teams already have governance — on paper. There is a refund cap, an approval matrix, a data-handling policy, a list of who can touch production. None of it reaches the agent making the call.

The rules live in one system, the decision happens in another, and the gap between them is filled with hope.

Prompt instructions drift. Wiki pages go stale the day they're written. Reviewers can't be in the loop for every tool call an agent will ever make.

You do not need more policies. You need the ones you already have to be reachable, in the hot path, at the moment of decision.

Govern is the act of turning stated rules into an artefact an agent can consult: a compiled, versioned bundle with a fingerprint you can point at. When the bundle changes, the fingerprint changes. When a decision is made, its fingerprint is stamped alongside the verdict.

Aarmos treats a policy bundle as a first-class object — authored, reviewed, signed, and versioned like code.

• Rules are written once and reused across agents.
• Each publish produces a fingerprint the runtime pins to.
• Old bundles never disappear; past decisions stay replayable.

A team ships a "no PII to third-party tools" rule. It compiles into bundle pol-2f9c… and is loaded by every agent in the workspace within seconds. Two months later an auditor asks which rule applied on a specific date — the fingerprint on that decision points back to the exact bundle, byte for byte.

Isn't this just a rules engine?

A rules engine evaluates rules. Governance decides which rules a given agent, in a given context, is bound by — and produces evidence that they were.

Do I have to rewrite everything as code?

No. Start with the two or three rules your team already argues about. The bundle grows as the questions do.

What about approvals that need a human?

Governance covers ask as a first-class verdict — a human can be the rule. See Authorities.

• Run
• Prove
• Authorities
• Receipts