[throwaway7783]

For sure. Not hard fails, but bad fixes. It confidently thought it fixed a bug, but it really didn't. I could only tell (it was fairly complex), because I tried reproducing it before/after. Ultimately I believe there was not sufficient context provided to it. It has certainly failed to do what I asked it to do in round 1, round 2, but eventually got it right (a rendering issue for a barcode designer).

These incidents have been less and less over the last year - switching it Opus made failure frequencies less. Same thing for code reviews. Most of it is fluff, but it does give useful feedback, if the instructions are good. For example, I asked for a blind code review of a PR ("Review this PR"), and it gave some generic commentary. I made the prompt more specific ("Follow the API changes across modules and see impact") - it found a serious bug.

The number of times I had to give up in frustration has been going down over the last one year. So I tend believe a swarm of agents could do a decent job of autonomous development/maintenance over the next few years.