Google: multi-agent team formalizes Weil conjectures in 1.8M lines of Lean via 28,800 subagents
rseroter · x · 2026-11-16
Google Research shares results from Antigravity's /teamwork multiagent systems, including a fully autonomous multi-agent team on Gemini 4 that formalized the Weil conjectures into 1.8 million lines of Lean code across 28,800 subagent invocations.
- Built on Gemini 4 Argon/Carbon/Hydrogen, the system targets long-horizon reasoning: agents write custom solvers, explore large search spaces, and verify long logic chains while human experts frame questions and constraints.
- Results span information theory, formal verification, molecular structure discoveries, and kernel optimization, plus production infrastructure gains inside Google.
- Two limits of current LLMs motivated /teamwork's design.
- The agentic tools (/teamwork, /boost) are now available externally, with a researcher access program.
More from coding & agent
- OpenTeams engineering blog: pip supply-chain incident, Polars 2.0, cross-page PDF OCR — KhuyenTran16 · 2026-11-16
- Dev cuts movie board with ~5,000 cards from 9s to under 3s load time — dSebastien · 2026-11-16
- Figma MCP server offers design file navigation and asset extraction with 95% token-optimized output — modelcontextprotocol · 2026-11-16
- Grok Bot 1.0 lands with bots that join meetings, take initiative and self-improve — mattyp · 2026-11-16
- Does grammar-constrained decoding hurt small or quantized models? A call for real numbers — bye_ella · 2026-11-16
- What's the Playwright equivalent for agentic desktop control? — jinnyjuice · 2026-11-16