Code Agent: delegate the change, review the diff

How it actually works

1. You describe the outcome — 'migrate the auth module to the new token format, keep the API stable, cover it with tests.'
2. The agent plans — a step list appears before any code changes: read, plan, edit, test, report. You can edit the plan or veto it.
3. Per-step model routing — architecture reasoning goes to Claude Opus or GPT-5; mechanical edits and test scaffolding go to DeepSeek; the final self-review goes to a different model than the one that wrote the code, because models catch each other's mistakes better than their own.
4. Sandboxed execution — edits land in an isolated clone with dependencies pre-installed. Tests run there. Your working tree is untouched until you accept.
5. Receipt — every step logs model, tokens, credits and wall-clock time. Total cost of the task is printed at the end.

What it's for — and what it isn't

Strong: migrations, refactors, dependency upgrades, test backfills, bug fixes with reproduction cases, API surface changes across many files. Not for: ambiguous product decisions, UI taste, anything where 'what should this even do' is the hard question. Give it work with a definition of done.

Keep reading

• Claude guide — the agent's main coding model
• DeepSeek — the cheap workhorse it routes to
• Research Agent