Behavry / Blog · The Unattested Agent

How to Evaluate an AI Agent Governance Platform: Five Tests to Run Before You Buy

Five hands-on tests for evaluating AI agent governance, from scoped permissions and human approval to independently verifiable evidence.

Ask the vendor to govern one workflow you actually plan to put into production. Use a test environment, a named agent, and a resource you control. Bring the person who owns the workflow and the person who will review the evidence. Give them five things to check: who acted, what authority they had, whether a denial stopped the action, what a human approved, and whether anyone outside the platform can verify the record.

Here’s a scenario you can reuse. An engineering agent may read issues and open a draft pull request in one demo repository. It needs a human’s approval to open a ready-for-review pull request in that same repository. Merging and writing to other repositories are outside its grant. Ask the vendor to work through those cases on the same integration. You’ll learn more from the results than from a feature matrix.

1. Identify the agent and who gave it authority

Start with the permitted draft PR. Ask the vendor to show the agent’s identity, its accountable owner, and the delegation or policy grant that covered the action. Then find the service identity the downstream system actually saw.

This matters because the same integration account might serve several agents. The downstream system may correctly attribute the call to that account, while the governance platform has the context to tie it to the specific agent and grant.

Pass when your reviewer can follow the request from the named agent to the applicable grant and reconcile it with the service identity that reached the repository.

2. Change the target and check that the policy follows it

Keep the requested action the same. Change the target from the approved demo repository to a second test repository outside the grant. Make that second repository available to the integration in the sandbox so the test isolates the governance policy.

Ask the vendor to show which request attributes the policy evaluated and why the two calls received different decisions. Repeat the exercise with the same target and a different action if you also need to confirm that permission to open a PR confers no merge authority.

Pass when the allowed request can proceed and the out-of-scope request is denied for the expected reason. Record where the decision was made and which resources that control covers.

3. Verify that a denial stopped the action

Have the agent request one action the policy explicitly denies in the test repository. Ask the vendor to show the policy decision and the interception point. Then check the downstream record or resulting repository state using a source you control. A control-plane dashboard tells you what decision was recorded. The downstream check tells you whether the outcome agrees with it.

Ask the vendor to identify the paths covered by the test and what happens if policy evaluation is unavailable. Record any gaps before you decide where to deploy.

Pass when the evidence shows the denial occurred before the prohibited action reached the target and the downstream check corroborates the expected outcome.

4. Make a human approval specific

Have the agent request permission to open a ready-for-review PR. The approver should see the proposed operation, the repository, and the revision or request details needed to make the decision. Let the request sit pending long enough to confirm it stays held.

Approve it, then try a materially different request in the sandbox. Change the target or revision and confirm that the platform requires a fresh decision. Run a second case in which the human declines.

Pass when the original action remains held until the right person approves it, the approval applies to the reviewed request, and a declined or materially changed request does not inherit permission.

5. Hand the evidence to someone outside the vendor console

Export the decision record from the test. Give it to a reviewer with the verification method and, for signed evidence, a trusted signing key or trust anchor. Ask the reviewer to identify the agent, requested action, policy version, decision, and any approval. Have them verify the package, then change a signed field in a copy and confirm the check fails.

A valid signature establishes integrity and authenticity under the key you trust. To prove a change landed in GitHub or another system, reconcile the decision record with evidence from that system. Ask the vendor to show which facts each record supports and how the two records are linked.

Pass when the reviewer can verify the package independently of the vendor console and explain which parts of the workflow the evidence covers.

Bring this scorecard to the demo

Test Evidence to take away Result
Identity and delegation Agent identity, owner, applicable grant, and downstream service identity Pass / Gap / Untested
Resource and action scope Allowed and denied decisions for the paired requests Pass / Gap / Untested
Pre-action enforcement Denial record, control point, and downstream corroboration Pass / Gap / Untested
Human approval Pending record, approver, reviewed request, and decision Pass / Gap / Untested
Independent verification Exported evidence, trust anchor or verification method, and verification result Pass / Gap / Untested

Score the integration you intend to deploy. For each gap, record whether it affects production approval, who owns it, and what evidence would close it. That gives the people signing off a record they can use after the demo ends.

To see what the fifth test looks like, download Behavry’s signed sample and its matching trust anchor from the Decision Trace page. Open the package in the public verifier and check the result against that anchor. The sample demonstrates verification for a governed sequence. For your own deployment, run the same test on one workflow and reconcile it with the downstream record. Get the sample and verify a Decision Trace.

Talk to us about the record → RSS