Show HN: Diff two versions of an AI agent and catch silent permission changes
A new tool lets you compare two versions of an AI agent structurally, detecting silent changes in tool permissions and trust levels without relying on an LLM. It runs entirely in the browser, ensuring privacy, and can be integrated into CI/CD pipelines.
Version A (before)
Version B (after)
Why this is deterministic, not another LLM
Ask a chatbot to compare two prompts and you get a different essay each time — an opinion, not an audit. This tool parses both versions structurally, derives each one's Trust Level with the same conservative logic as the AgentAz validator, and diffs the tool boundaries directly. Every finding cites the exact change (“issue_refund lost its approval gate”), and the radar shows the shape of the risk surface before and after. Increasingly, the thing that rewrites your agent is an LLM — which is exactly why the referee shouldn't be one.
Runs entirely in your browser: nothing you paste is uploaded or stored. To catch drift continuously instead of manually, track the agent in Agent Watch or enforce a maximum tier in CI with the GitHub Action.