AI agent receipts: what to record and how to check them

An AI agent receipt is a record of a specific tool execution. It helps a person or another AI inspect what was requested, what ran, and what the system observed without relying on the agent's final chat message. A useful receipt has a stable identifier, tool name, timestamp, status, observed result, and material for an integrity check.

Start with the execution boundary

Separate the agent's intention from the tool's result. A plan is not an execution, and a successful local tool response is not necessarily a completed action in another system. For a state-changing task, preserve the original response and seek a callback, transaction, or read-back from the system that owns the state. If that confirmation is missing, label the outside outcome as unconfirmed.

Check a receipt step by step

  1. Open the public receipt page and note its ID, tool, timestamp, status, and result preview.
  2. Call the verifier independently and confirm that its ID matches the page you inspected.
  3. Check the verifier's integrity result and output hash. When canonical bytes are available, recompute the digest using the documented serialization rule.
  4. Inspect provenance and external evidence separately. Follow an upstream reference when the task claims a payment, delivery, update, or other outside effect.
  5. Report exactly what was verified and what remains unknown. A missing evidence field must not be rewritten as success.

For a live example, open this public execution receipt and request its verification result. You can make the same request from a terminal:

curl -fsSL https://zambo.dev/api/receipt/83ae4fd7-b9f8-497c-b842-72381c8ee6a8/verify

What an integrity check can and cannot establish

A passing verifier result checks the stored record according to the verifier's rules. It can help detect changes to the recorded bytes and make the observed execution reviewable. It cannot independently prove that an external provider completed an action unless that provider's actual confirmation was captured and bound to the record. A screenshot or an agent summary is not a substitute for the original record and its verification path.

Use receipts as a handoff

Save the public receipt URL with the task status when work passes from one AI client or reviewer to another. Keep secrets, private prompts, and raw credentials out of the public projection. If a call failed, was blocked, or is still pending, preserve that state so the next person can choose a safe next step instead of repeating a consequential action.