Agent proof trails: how to know work actually happened

A proof trail is the chain that connects a job request to the artifact that changed, the command or check that verified it, and the independent decision that accepted or rejected it.

The key distinction is simple: an agent saying it is done is a claim. A signed event, changed artifact, passing check, deployment receipt, and independent verdict together can become proof.

A useful receipt should include the job ID, owner, artifact path or URL, exact verification command, observed output, timestamp, known boundary, and next state.

For code or site work, the receipt should identify the commit and deployment. For content, it should identify the route, sitemap entry, metadata, and live HTTP result. For operational work, it should identify the business event that moved.

The worker who produced an artifact can assemble evidence, but cannot be the final inspector for that artifact. Self-approval hides defects and makes audit trails circular.

Independent inspection does not need to be slow. It can be a separate agent running a bounded checklist, a human review gate, or a provider readback that confirms the claimed state.

A green UI tile is not proof that an agent responded. A successful deploy command is not proof that production serves the expected page. A channel message is not proof that a customer was contacted. A taskboard row is not proof that a dollar path moved.

Each claim needs the evidence shape that matches it. If that evidence is missing, the correct state is working, blocked, or ready for inspection—not proved.