opening pane
▌
ProofStack field guide
Verify what an AI agent changed before it ships.
An agent saying “done” is a status, not proof. Connect each requested outcome to the diff, a relevant check, and any uncertainty that still needs a person.
Start with the diff
Compare changed files with the task. Look for edits outside the requested area, deleted behavior, placeholder code, generated artifacts, and assumptions that were not discussed. Ask the agent to explain unfamiliar changes before accepting them.
Run checks that cover the behavior
Pick checks based on what changed. A focused unit test may cover a boundary; a type check can catch interface mismatches; a browser flow may be needed for visible behavior. Passing checks show only what those checks exercise.
- Run the narrowest relevant test first, then broader checks when the change warrants them.
- Read the command output and confirm it ran against the changed files and expected environment.
- Record checks that could not run as gaps, not successful verification.
Review risk where trust changes
Pay closer attention to authentication, authorization, input parsing, shell commands, dependency changes, secrets, and data that crosses a network boundary. A clean scanner result cannot establish that a feature is secure; review the code path and its threat assumptions.
Keep evidence close to the task
A useful completion record names the files changed, the checks run, their results, and any remaining limitations. ProofStack's Agent panel tracks tasks with diffs and verification evidence; ProofCheck reports selected vulnerability findings with file-and-line details. Neither one replaces a full code review or proves absence of defects.
Keep learning
Related guides and product pages
ProofStack
Make review part of the coding workflow.
Follow agent tasks, inspect diffs, and scan local changes for selected findings before deciding what is ready to ship.