Use an agent without outsourcing understanding

State the observable behavior, constraints, repository scope and a concrete check. Ask the tool to inspect the relevant implementation before editing, and give a small example of expected and failing behavior.

Keep the change reviewable: one coherent task, explicit dependency policy and the actual test command. An agent can propose code and run tools, but you still need to understand invariants, failure modes and the diff.

Evaluate evidence: inspect changed files, run the meaningful tests/build and review error handling and ownership. Product features and defaults evolve; check the official documentation for the installed tool instead of copying stale flags or assuming every agent has the same permissions.