Google details /teamwork: long-horizon agents behind its research breakthroughs
ymatias · x · 2026-11-14
Google Fellow Vahab Mirrokni detailed how Antigravity's /teamwork multiagent system (built on Gemini 4 Argon/Carbon/Hydrogen) produced recent breakthroughs: a full proof of the Courtade–Kumar conjecture with Lean 4 verification, 95% top-1 NMR-based molecular structure determination, automated safe Rust kernel rewrites, 4x TPU kernel speedups, and 1+ PiB fleet memory savings.
The design targets two LLM limits — bounded generation length and error-prone long reasoning chains — via an interactive discovery loop where researchers frame questions and constraints while long-horizon agents write solvers, explore large search spaces, and verify each step. Tools are open in Antigravity, with 6 months free access for 10,000 academic researchers.
More from coding & agent
- Free metroidvania level generator turns a seed into a full tiled world — majidmanzarpour · 2026-11-14
- Open-source Agent Skill wraps DeepSeek API for docs, OCR and TTS — vista8 · 2026-11-14
- Fable-style scene rebuilt in pure three.js code, zero assets used — vinis_artstreaks · 2026-11-14
- 5 prompts this developer uses daily with Claude to get better coding results — jarrodwatts · 2026-11-14
- AI-designed main menu concept art for 'The Rat: Director's Cut' — chrisfirst · 2026-11-14
- agent-browser v0.43 cuts install size to 17MB and startup to 2ms — cramforce · 2026-11-14