Agent Action Governor
A policy layer that decides before an agent acts — and records every decision so you can prove it.
The problem
/ 01Autonomous agents act fast and at scale, and a single bad action can be costly. Teams need a policy layer that decides up front and records every decision.
What it does
/ 04Defines allow / deny / escalate rules per agent action type.
Evaluates agent actions against the policy with a clear matched rule.
Approves escalated actions with an audit note.
Lists the full evaluation and approval audit trail.
Private pilot adds
/ pilotPilot-ready: durable agent-governor store available under NDA with a /v1 API for private pilots. The public studio card runs against /demo (ephemeral).
Who it’s for
/ 03- Teams deploying autonomous agents
- Security and governance teams reviewing agent actions
- Platforms that must prove agent behavior stayed within policy
In context
/ 01Evaluations feed the AI Decision Audit Trail — policy decisions and the approvals that follow them land in the same record.
Spec & limits
/ limits| Interface | Input limits | Output artifact | Failure mode | Non-claim |
|---|---|---|---|---|
Studio ephemeral /demo; pilot durable /v1 on your infrastructure. | TBD — verified in pilot acceptance | TBD — verified in pilot acceptance | TBD — verified in pilot acceptance | TBD — verified in pilot acceptance |
Related guide
/ guideSee the agent policy gates guide for a deeper problem walkthrough and verification notes — evaluation notes for the live demo and for a private pilot. Also: try the live experiment and read the product overview.
Frequently asked questions
/ 0501 What does "allow / deny / escalate" map to at runtime?
Before an agent acts, the gate evaluates the action against signed policy: allow passes it with a receipt, deny blocks it with an audit entry, and escalate routes to a human/approval step.
02 Are policies signed so the agent cannot tamper with them?
Yes — policies are signed and versioned; the governor verifies the signature before enforcement and emits a receipt binding the action to that policy version.
03 How is each allowed (or blocked) action recorded?
Every decision emits a hash-chained, signed receipt — allow, deny, and escalate alike — so approvals and blocks are equally auditable.
04 Can policies be updated while agents are running?
Yes — a new signed policy version supersedes the prior one; in-flight actions finish under the old version and new actions evaluate against the updated policy.
05 Does this replace the need for prompt-level guardrails?
No — it is a control layer that sits above prompts, not a replacement for prompt engineering or model-level guardrails.