Agents patch symptoms until you tell them not to
2 min read
Across three PRs, the agent handled review comments one at a time until I explicitly asked it to connect the failures and investigate the underlying mechanism.
I saw the same review loop on three pull requests, weeks apart. A bot left a comment, the coding agent made the smallest plausible fix, and the bot returned with another version of the same complaint. I initially blamed the agent. After the third PR, it was harder to ignore the workflow I had put it in.
Each review comment points to a local line, so a local edit is a reasonable response. The loop never asks whether comments three, seven, and eleven come from the same broken assumption. It asks for the current comment to be resolved, and that is what the agent does.
On one of those PRs, I stopped asking for another fix and named the pattern:
look for the root of all these issues... instead of going back and forth like this and just doing surface level bandaid fixes
That changed the response from another patch to a diagnosis:
You're right — I've been patching symptoms... That's undecidable by text-scanning SQL. Every Codex round is a new manifestation of the same root error.
The check was trying to prove a runtime property by scanning SQL text. Each false positive looked different at the line level, but all of them came from that design mistake. Once the agent examined the mechanism, it removed the fragile linter and replaced it with a deterministic sweep.
There are real limits to this observation. All three examples came from one repository, and I used nearly the same correction each time. The bot-driven review loop may itself encourage comment-by-comment work, and the agent may have learned to wait for my familiar wording. Three occurrences are enough for me to change my process, not enough to claim that every agent behaves this way.
A simple comparison is enough to test the idea. Take a pull request on its second or later review round. In one fresh session, ask the agent to resolve the latest comment. In another, ask it to connect the latest comment to the earlier ones and investigate the root cause before editing. Compare the investigation and the diff. If both runs behave the same, that is useful evidence too.