Are an agent's own logs enough?

They can be useful for debugging and understanding what the agent reports, but by themselves they are self-reported. They do not independently establish that a tool ran or that an external change succeeded. Compare them with execution records or the affected system.

What does a third-party-verifiable receipt prove?

A receipt that another person can inspect can support checking what the execution system recorded, such as a tool call, its observed result, and integrity information. It does not automatically prove that the result was correct or that an outside system completed an outcome unless that evidence is included and checked.

When should a person review the work?

Use human review when the result is consequential, ambiguous, subjective, or difficult to confirm automatically. A reviewer can inspect the relevant source system and judge whether the result meets the request; this complements rather than replaces a reliable execution record.

Does a Zambo receipt prove an agent completed a task correctly?

A Zambo receipt provides a way to inspect and verify recorded tool execution. It can help establish what was recorded, but it does not by itself certify correctness or an external outcome. Check the resulting state or independent confirmation when that matters.

How Do I Prove My Agent Did the Work?

Q: How do I prove my agent did the work?

Match the agent's claims to evidence from the system that performed the work. Use logs to understand the sequence, an independently checkable execution receipt to inspect what a tool recorded, and a human or source-system check for important outcomes. The right evidence depends on the stakes.

Last updated: 2026-09-27

Match what the agent says it did to evidence from the system that performed the work. Logs explain the claimed sequence, verifiable receipts let others inspect recorded execution, and human review checks whether the result actually meets your needs. No single method proves every kind of task.

Try the Zambo demo · Explore Zambo

Which evidence should I use?

Are self-reported logs useful?

Yes. Logs and transcripts can help reconstruct the sequence, troubleshoot failures, and compare the agent's account with later evidence. Their limitation is provenance: if the agent or its host produced the log, it is still a report from that environment, not independent confirmation that an action ran or changed anything.

What do third-party-verifiable receipts add?

A receipt that another person can inspect gives a reviewer a record tied to an execution, such as the tool call, observed result, time, and integrity material. It can make the execution record independently checkable without relying on the original chat. A valid receipt is not a guarantee that the tool's result was correct, nor proof of an outside outcome that it did not observe.

What does human review add?

A person can compare the result with the original request, inspect the affected file or system, and judge quality or context that a record cannot assess. Human review takes time and can make mistakes, so for consequential work it is often strongest when guided by an execution record and direct source-system evidence.

How do I prove my agent did the work?

What is a practical verification process?

Write down the important claims in the agent's completion message. For each, ask what system would show the action or outcome. Compare the agent's log with an execution record where available, then inspect the changed state or seek a human review for anything important, subjective, or hard to reproduce. Keep execution evidence separate from evidence that the outcome was successful.

Is a receipt always better than a log or review?

No. A log may be enough to understand a low-risk process; a receipt can make selected executions easier to inspect independently; and a person may be needed to judge correctness or intent. Choose evidence proportionate to the stakes, and combine methods when a mistaken completion would matter.

How does Zambo fit into these options?

Zambo provides verifiable receipts for recorded tool executions and a public receipt verifier. That is one way to make execution evidence inspectable by someone beyond the original chat. It does not replace logs, human judgment, or checking an external system when you need to establish that an outcome actually occurred.