Step 1: Declare

Govern AI agents before they act.

Agent Registry turns agent permissions and policies into reviewable YAML contracts. Its runtime checker evaluates action and resource rules, signed bundles, and supported approval, evidence and failure-threshold policies before a wrapped call runs. Other contracts, including budgets and rate limits, need separate enforcement.

pip install hlinor-registry - Python 3.10-3.13. See the versioned tests and validation workflow.

The pain it solves

Your agent can call tools, spend API budget and touch customer data. Today that boundary lives in prompts and good intentions. When something goes wrong, nobody can show what the agent was allowed to do - or prove it stayed inside.

Common triggers

  • An agent with tool access handles PII or payments.
  • API costs spike and nobody can say which agent did it.
  • A client or auditor asks "what can this agent do?" and the answer is a shrug.
  • A model update quietly changes agent behavior in production.

How it works

Four steps from intent to enforced boundary.

Write the boundary

Declare skills, policies, allowed and blocked actions in a versioned YAML file your whole team can review in a pull request.

Compile it

The CLI compiles YAML into a signed, deterministic bundle. Static linting catches contradictions before deployment.

Enforce at runtime

Wrap existing tools with a decorator or framework integration. A denied action simply does not execute and returns a machine-readable reason code.

Export the evidence

Every allow/deny decision lands in a JSONL log, ready for Splunk, Datadog or your auditor.

What teams use it for

01

Prevent PII leaks

Block external email, record mutation and other exfiltration paths for agents handling sensitive data.

02

Make cost limits explicit

Declare budgets and rate limits beside the agent definition so reviewers can inspect them. Enforce those limits in the calling application or gateway; the current PolicyChecker does not meter spend or rate-limit calls.

03

Answer procurement

Hand enterprise buyers a signed policy bundle and decision log instead of a promises document.

Next step

Declared is not verified.

Agent Control Plane scans your actual codebase and runtime to check that what is deployed matches what you declared.

Agent Control Plane

Want this reviewed independently?

Our Agent Governance Review checks your agent policies, runtime permissions and audit trails - and delivers a client-ready report in 72 hours.

Agent Governance Review