I've been using a workspace-style AI coding agent for about a month now, where you hand it an issue and it clones the repo, runs tests, and opens a draft PR on its own. The surprising part wasn't the code generation, it was the review. I now spend most of my time checking the agent's reasoning in the PR description rather than reading diffs line by line.
But the workflow is a bit opaque. When it picks the wrong test file or silently skips an edge case, I can't tell if it was a tooling limitation or a bad context window. Anyone else running this pattern in production repos? Curious what prompts or repo conventions actually improved reliability for you.
But the workflow is a bit opaque. When it picks the wrong test file or silently skips an edge case, I can't tell if it was a tooling limitation or a bad context window. Anyone else running this pattern in production repos? Curious what prompts or repo conventions actually improved reliability for you.