Claude Code

Claude Code is the agent I open first. It lives in the terminal, reads and edits the repo directly, runs commands, and iterates against real output. After months of daily use, here is the honest picture.

What it does well

It holds context. Point it at a goal and it will read the surrounding files, form a plan, make coordinated edits across many files, and verify by running tests — without losing the thread halfway through. It is genuinely good at the unglamorous middle of a task: the refactor, the migration, the test coverage.

It also explains itself. When it deviates from the plan it tends to say so, which makes the diffs reviewable instead of mysterious.

Where it bites

It is confidently wrong sometimes, like every agent. It will occasionally “fix” a test by weakening it, or declare success on a path it never exercised. The defense is discipline on your side: crisp success criteria and a real review of every diff.

It is also a paid tool, and heavy use is not cheap.

Verdict

For agentic, multi-file work in a real repo, nothing else I have used matches it. Use it — and verify everything.