Welcome back to CodeCast. I’m your host, and today we’re diving into Martin Fowler’s exploration of autonomous coding agents. He starts by distinguishing supervised chat tools like GitHub Copilot from true “autonomous background coding agents” such as OpenAI Codex, which spin up their own environment and return a pull request.

Fowler gave Codex a simple cosmetics task: “Improve the sophistication of the ‘category-to-human-readable’ logic so that labels appear as ‘Client Research’ and ‘Delivery Management.’” The log reads like a terminal play-by-play: endless grep searches, file edits, test failures, and environment setup frustrations. He notes how Codex resorts to “brute text search” instead of semantic indexing, and that “the remote dev environment is key” for reliable results.

Finally, he measures solution quality across six runs, finding that only a third of them reused existing code or caught all edge cases. Fowler leaves us with a provocative question: when these background agents deliver 80 percent of the fix, who cleans up the last 20 percent?
Link to Article