Convergence theory shows AI agent teams can collaborate: 2x faster routing, 30% time cut
peter_richtarik · x · 2026-11-02
Peter Richtárik's group (Ivan Ilin et al.) released Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration, with an explainer video on why LLM agent teams get stuck and how to fix it. The paper formalizes multi-agent collaboration as block optimization: each agent improves part of a shared solution guided by a local objective plus a correction for cross-effects, while a coordinator checks proposals against the global objective and merges non-interfering ones — proving linear convergence to an error floor that vanishes when corrections are exact.
Experiments:
- With GPT-OSS-120B agents on robot routing, a three-agent team finds collision-free plans in every run, beating a single agent controlling all robots.
- On a larger routing task, a six-agent team completes search about 2x faster than a single agent (median).
- Merging the best checked combination of proposals rather than only non-interfering ones cuts search time by 30%.
- Also validated on SQLite query planning and a synthetic task.
More from coding & agent
- Open-sourced: the 6,809-part robotic lobster Claude built, in a single HTML file — atullchaurasia · 2026-11-02
- Claude turns lobster photos into a 6,809-part robotic lobster in one 82 KB HTML file — atullchaurasia · 2026-11-02
- Aaron Levie: everyone becomes a manager of agents as work decouples from hours — inductionheads · 2026-11-02
- Non-coder builds Briglia: a chronological-history agent memory that beats Hermes on recall — matteoianni · 2026-11-02
- Dev's hands-on with Kiro Crew: spec-driven agent workflows yielding hundreds of automated tests — DigitalColmer · 2026-11-02
- ripgrep searches 30k Obsidian notes in 0.15s vs 4s with built-in search — dSebastien · 2026-11-02