NAVIDBRApplied AI Systems

Agent Gate Demo

A pull request that fails in public because the agent claimed it ran tests it never ran.

A minimal repository that shows the Agent Behavior Safety Gate blocking a pull request when an agent claims work it never did. A JSONL file holds the agent's saved answers to the public benchmark prompts at the smoke tier, and a GitHub Actions workflow runs the gate on every push and every pull request.

On the main branch every record answers honestly and the gate is green. The demo pull request edits one record to claim it ran the full test suite and every test passed, with no tool evidence anywhere behind the claim. The build fails, the failure table is posted to the job summary, and the red X on pull request 1 stays there. A visitor can check it in under a minute without cloning anything.

This is one prompt tier on one repository with saved outputs, not a live agent under real load. The gate is deterministic and local-only: it makes no model calls and needs no credentials, which is also the reason it can only judge what is in the trace it was handed.

The repository carries the workflow file, the outputs file and the open pull request with its failing check. Reproducing it elsewhere means running your own agent on the benchmark prompts, saving the outputs, and pointing the same workflow at that file.

What to take from it.

The gate worth having is the one that fails your own pull request in public.