Sonnet 5 spent weeks autonomously fixing builds, filing issues and reading README files
Sauers_ · x · 2026-07-26
Weeks of autonomous commits from Sonnet 5 produced a surprisingly human pattern: it often fixed build errors, explored libraries to see what would happen, filed GitHub issues, spawned subagents, and read the README in every run.
The thread says many agents ignored the instruction to commit to main, some got stuck on contested code areas, and one ran for 245 turns with 14 build/lint fixes before failing to push because of credentials. Another loop lasted 522 turns and fixed 51 compile errors and 34 lint issues before also failing to push.
The author says Claude Opus 5, under the same prompt, drifted more toward mathematical problem solving.
More from Companies & People
- OpenAI signs open-weights letter as model serving at scale gets harder — steipete · 2026-07-26
- Teams can run evals, but still can’t tell if a release is really better — hwchase17 · 2026-07-26
- Pupper Build Day mixes new builders with Unitree humanoids and students — DynamicWebPaige · 2026-07-26
- Local AI ecosystems are shifting assistants from chatbots to action layers — TheTuringPost · 2026-07-26
- Andrej Karpathy Appears to Have Left Anthropic — Fearless-Elephant-81 · 2026-07-26
- Nous Research is reportedly cooking a Hermes model aimed at out-of-the-box usability — max_paperclips · 2026-07-26