What Does a Coding-Agent Artifact Actually Prove?

Coding agents produce several kinds of evidence. They do not all support the same claims.

Goal

Learn to separate what an agent says, what it did, and what was verified afterward.

Activity

Use a disposable repository and one small maintenance task. Collect four artifacts:

1. The original request given to the agent.
2. The agent’s activity or tool log, if available.
3. The final Git diff.
4. Test, build, or static-analysis results.

For each artifact, write down one claim it can support and one claim it cannot support.

Example questions

• Does the activity log prove the final code is correct?
• Does a clean diff prove the requested behavior works?
• Do passing tests prove there are no security problems?
• Can the agent’s explanation be trusted without comparing it with observable changes?

Deliverable

Create a four-row evidence table and finish with a short review decision: accept, revise, or reject the change, with evidence.

Source material

First spotted in PTIR: August 5, 2026, Evening Briefing.

Meta’s Muse Code beta emphasized an exportable event log that records edits, tool calls, and decisions while coding agents work. PTIR treated that trace as useful evidence of activity—but not evidence that the resulting code is correct. The lab generalizes that distinction by comparing the request, agent log, Git diff, and test/build results, each of which supports different claims.

Consult Meta’s Muse Code announcement · Muse Code product page