# How do I know my AI agent finished the work, and get proof?

_Verify agent completion with an observable definition of done, linked artifacts, passing checks and live evidence instead of a chat claim._

By Saad Ahmed, 2026-10-20

To know whether an AI agent finished, compare the delivered result with a written definition of done and inspect the evidence attached to each requirement. Treat “done” in chat as a claim until tests, artifacts, review and live behavior support it.

The proof should match the work. A commit can prove that code changed; it cannot by itself prove that the feature works in production.

## Define observable completion

Write acceptance criteria before execution. Each criterion should describe something a reviewer can observe, such as:

- a named test passes;
- a page returns 200 and shows the intended content;
- a pull request is reviewed and merged;
- a deployment contains the tested commit;
- a document answers the required questions;
- a screenshot, trace or recording demonstrates the interaction.

Avoid criteria such as “implementation complete” or “works well.” They repeat the agent's conclusion instead of defining evidence.

## Build an evidence packet

Ask the agent to leave a compact packet with the result link, commit or version, commands run, relevant outputs, known limitations and any check that still needs a person. The packet should let a fresh reviewer verify the outcome without reconstructing the session.

Use independent signals where risk is higher. For a web change, combine automated tests with a live URL check. For data migration, keep counts and reconciliation output. For a document, verify both source and rendered result.

## A repository workflow

Put the definition of done in the issue or project file. Require the agent to link its branch and pull request, then report exact checks. A person or review agent evaluates the evidence before changing the status to done.

This works without a dedicated product, provided the evidence stays attached to the task rather than only in chat.

## Where ecco fits

ecco keeps the outcome, proof requirements, agent updates and artifact links on one task. Connected agents can attach commits, pull requests and test evidence; the board can then separate work awaiting review from work accepted as done.

[Connect an agent to ecco](/blog/connect-an-agent). For why a shared record differs from relying on one assistant's conversation, see [ecco versus Claude and ChatGPT](/blog/ecco-vs-claude-and-chatgpt).

## Frequently asked questions

### Is a passing test suite enough proof?

Only if the definition of done is completely covered by those tests. Deployment, content quality, accessibility or a human business decision may require additional evidence.

### Who should mark an agent task done?

Use your risk level. Low-risk mechanical tasks may close automatically after required checks; consequential work should wait for a person or designated reviewer to accept the evidence.

### What if the agent cannot run a required check?

It should say which check was not run, why, and what a reviewer must do next. The honest status is awaiting verification or blocked, not done.

### Should I keep full agent transcripts as proof?

Transcripts can help an investigation, but they are not a concise completion record. Preserve the final artifacts, relevant checks, decisions and limitations on the task.