Your coding agent says it fixed the bug. That's a claim, not a result.

Agents are pretty good at fixing integration bugs now. They're even better at sounding like they did.

The failure mode I kept hitting: a confident summary of a fix that didn't happen. Or a fix that made the symptom disappear while the data stayed wrong. Swallow the exception, error goes away, test goes green, customer still can't access what they paid for.

So we stopped taking the agent's word for it. Three things now get verified server side:

- Did the app actually run in both states (buggy and fixed)?

- Did the exit code flip from failing to passing?

- Did the side effect itself change: the balance, the entitlement, the row that was wrong?

An agent narrating confidently can't produce those. Either the run happened or it didn't. Anything short gets marked self-reported and the gate refuses to certify. It's annoying to have your own tool call a fix unproven. Much less annoying than shipping it.

We built this as an MCP server so the agent already in your editor can reproduce, fix, and prove without leaving the IDE. Works against 70+ APIs, Stripe, Clerk, HubSpot, RevenueCat and so on.

How are you verifying an agent fix actually changed the right thing, not just silenced the error?