Atlas · agents
Verification
Checking the result before claiming success.
See it working: Plate I
In plain words
“Done” from a model only means it predicted the word “done”. A good agent checks its work first, the way Plate I’s agent reads back its edit.
How it works
Verification can be as small as re-reading a changed file or as large as running a test suite. Claude Code describes verifying as part of its normal loop.
Going deeper
Across many runs, repeatable evaluations compare versions of an agent workflow, so you can tell whether a change to instructions or tools actually helped.
Sources
- How Claude Code works Anthropic
**gather context**, **take action**, and **verify results**
- Agent evals OpenAI
repeatable datasets and eval runs.
Quotes were checked against the live pages on 2026-11-08.