Trust but Verify Your AI Agent

Run

Give the agent a bounded request with clear permissions and a defined stopping point.

Receipt

Keep the public record for each consequential tool call.

Check

Use the verifier and compare the result with the request and expected state.

Continue

If evidence is missing or the call is blocked, ask for a safe next step instead of silently assuming completion.

Frequently asked questions

Who authored the AER-1 draft?

trust but verify AI agent

Trust but verify means allowing an AI agent to work while requiring a checkable execution record before accepting a consequential claim. Run the bounded task, open the receipt, inspect the observed result, and continue only when the evidence supports the next step.

What can a verifiable receipt prove?

It can show what the execution layer recorded and what its integrity checks verify. It cannot prove an unobserved outside outcome.

How can another reviewer check the record?

Open the public receipt, compare the tool, time, result, and status, then call the verifier. Keep outside confirmations separate from the execution record.

What this question can establish

The practical question behind Trust but Verify Your AI Agent is whether the record supports the claim being reviewed. A user wants an agent to work within a bounded permission set while preserving a check before consequential claims are accepted. Separate the request, the execution, the system's observed output, and confirmation from any outside system. These are different evidence boundaries, and a reviewer should not treat them as interchangeable.

Use risk-based checks: require stronger independent confirmation as the consequence of an action increases, and ask for approval when the boundary is unclear. A useful conclusion names the evidence that was checked and the conditions that remain unresolved. This keeps an execution record useful without stretching it into a broader assurance claim.

Evidence to retain for this workflow

Start with the smallest set of fields that lets a second reviewer identify the event and repeat the relevant check. For this topic, Define the requested state, permitted tools, approval boundary, expected read-back, stop condition, and receipt needed for each consequential step.

• Task boundary: record the requested state and what would count as acceptance.
• Execution identity: preserve the receipt ID, tool name, caller scope, time, and status.
• Committed result: keep canonical bytes or the published representation needed to recompute the output hash.
• Provenance: distinguish an action the system ran from one it observed or merely received as a report.
• External confirmation: capture a real callback, transaction reference, read-back, or provider record when the task depends on outside state.
• Review context: note who checked the record, which verifier was used, and any limitation that affected the conclusion.

A repeatable review sequence

Use the same order for routine checks and incident review. It prevents a later summary from replacing the evidence the task actually requires.

1. State the requirement. Write the expected outcome and the exact source that can confirm it. A user wants an agent to work within a bounded permission set while preserving a check before consequential claims are accepted.
2. Locate the execution. Match the public receipt identifier to the requested event, then confirm the tool, scope, timestamp, and status.
3. Check integrity. Decode the committed bytes as specified by the receipt format, recompute fingerprint, and compare the complete digest rather than a shortened preview.
4. Check the owning system. When the claim involves an outside change, query that system or inspect its genuine callback. A receipt alone cannot supply a missing confirmation.
5. Write a bounded finding. Record what ran, what was observed, whether the bytes matched, and what remains unknown. Use risk-based checks: require stronger independent confirmation as the consequence of an action increases, and ask for approval when the boundary is unclear.

Failure states and misleading success signals

Preserve failed, blocked, pending, retried, and partially completed states as distinct outcomes. A plausible final message can hide a blocked call, partial result, or missing provider confirmation. Stop at the strongest evidence actually available. A successful transport response can still contain a tool error, and a valid integrity check can still protect an incorrect or incomplete result.

When evidence conflicts, keep both records and identify which system owns each fact. Do not overwrite an earlier receipt or silently promote a retry into proof that the original step succeeded. Explain whether the conflict is a timing issue, a stale read, a missing callback, a policy boundary, or a different task interpretation.

Privacy, retention, and safe sharing

Reviewability does not require publishing every input. Share a minimized receipt and safe task context. Do not widen public access to a full conversation just to make a single action reviewable. Prefer stable references and cryptographic commitments where they are enough for the reviewer, and keep the underlying sensitive record with the system that controls it.

Before sharing, check the public page, result preview, metadata, and linked evidence for secrets or personal information. Set retention and access according to the review purpose. If a record must be corrected, publish a new reference and preserve the original history rather than rewriting a record that another reviewer may already have captured.

Example of a careful conclusion

This kind of note gives the next person a useful handoff: what was checked, what the check means, and where the evidence stops. It is more informative than a bare success label and avoids presenting an unobserved event as a verified fact.

Common review questions

What does the record establish?

It establishes what the execution system recorded and whether the exposed canonical bytes match the output commitment. For trust but verify your ai agent, it does not by itself establish correctness or a result outside that system.

What if a provider or callback is unavailable?

Keep the state pending or unavailable, retain the error and time, and avoid claiming completion. In this workflow, A plausible final message can hide a blocked call, partial result, or missing provider confirmation. Stop at the strongest evidence actually available.

What should a reviewer inspect first?

Define the requested state, permitted tools, approval boundary, expected read-back, stop condition, and receipt needed for each consequential step. Then compare the record with the task requirement and ask the system that owns any claimed external state for a separate read-back.

What is safe to share?

Share the public receipt URL, identifier, status, verifier result, and a short sanitized task description. Share a minimized receipt and safe task context. Do not widen public access to a full conversation just to make a single action reviewable.

Further reading

• Reviewing agent work
• Execution evidence boundaries
• Task completion checks
• Agent handoffs
• Diagnosing incomplete work
• receipt definition