Is anyone getting genuinely good code out of coding agents?
Reviewing agent-written merge requests takes far longer than reviewing a colleague's, and the code is often convoluted and full of traps. Has anyone found a way to get clean, reviewable code from coding agents instead of just giving up on reading it?
Most hard-to-review agent code comes from one pattern: a vague request, one big run, and one huge diff at the end. The agent fills every gap with guesses, and you inherit all of them at once.
What helps is moving the thinking before the code. Ask for a short written plan first, cut it into small steps, and approve the plan. Each step then produces a diff you can read in a few minutes, and you can reject a bad direction before it spreads.
Tests are the second lever. If the agent has to write a failing test first and then make it pass, the change has a clear contract, and you review behaviour instead of guessing intent from the code.
Finally, make review a step the agent runs, not only you. A separate review pass with its own checklist catches dead code, duplicated logic and missing edge cases before the change reaches a human.
Listings mentioned
- Writing plans · skill by obraTurns a request into a small, step-by-step plan before any code is written.
- Test-Driven Development (TDD) · skill by obraMakes the agent write a failing test first, so every change has a clear contract.
- Requesting code review · skill by obraAdds a review pass against the requirements before the change is called done.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.