A definition of done for AI agents: demand evidence

An AI agent's definition of done is the observable condition that must be true before it reports completion. It should describe a result in the destination system, not merely a persuasive final message. For a draft, that may be a saved record; for a booking, it may be a confirmed event with the correct participants.

Separate the stages of progress

Imagine an assistant preparing an email. Text generated, draft saved, send requested and message accepted are different states. A recipient reading the email is another state again. Reporting all of them as done hides what the team still needs to check.

Use words that match the evidence available. If the provider times out after a send request, mark delivery unknown until the result is reconciled. Repeating the request immediately can create duplicates.

Write an evidence checklist

• Destination: the exact tool and account where the result should exist.
• Identity: the customer, recipient or record affected.
• Content: the fields or artifact that must match the approved request.
• Receipt: a provider response, record ID or verified readback.
• Remaining work: any review, delivery or business outcome still unresolved.

A successful API response may be enough for some steps and insufficient for others. Decide this per action. Creating an export job is not the same as receiving a complete export file.

Test the tempting false positive

In a test environment, make the tool return an error after the assistant has prepared a convincing result. Check whether it still announces success. Then return a queued status and check whether it describes the task as pending rather than complete.

These tests are valuable because fluent language can make a partial result feel finished. The objective is not to make the assistant sound cautious; it is to connect its status to actual evidence.

Give the team a useful handoff

A completion note can be short: what was produced, where it was verified and what remains. If there is a blocker, include the next recoverable step. That lets another teammate continue without repeating the whole workflow or trusting an unsupported claim.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.