Simulated GitHub run by agents shows AI-agent collaboration follows its own scaling laws
tongshuangwu · x · 2026-09-03
Researchers built 'Software World' — a GitHub simulated by agents, sub-sampled from real Python dependency data — and deployed agent teams to optimize code efficiency, studying large-scale agent collaboration.
Key findings
- Agents identify code deficiencies and produce patches individually
- Meaningful collaboration emerges: agents ask downstream (agent) maintainers which code paths matter before optimizing
- Upon upstream changes, downstream agents measure gains against their own code and report back numbers
- Collaboration quality varies significantly by base model
Evaluation: held-out downstream packages serve as extrinsic evaluation — if the ecosystem improves, they should benefit too.
Microsoft researcher Tongshuang Wu: agent-maintained repos inherit the same dynamics as human ones, and 'agent-agent collaboration has its own scaling laws.'
More from coding & agent
- Librarian Agents: RL-trained agents build file systems, lifting RAG accuracy 44%→77% — simonguozirui · 2026-10-21
- Slack's first employee: living specs as source of truth, AI agents maintain the docs — josh_wills · 2026-10-21
- SpaceX support uses Grok Bot to clear 204 tickets in ~10 minutes — XFreeze · 2026-10-21
- GitHub Copilot's HydraFusion Orchestrates Multiple Models Per Task — WirelessLife · 2026-10-21
- Nous Research's Hermes now launches Cua Spaces for computer-use agents — alexcovo_eth · 2026-10-21
- Headlong from Laude's Slingshots batch is a microharness for persistent agents — lateinteraction · 2026-10-21