Google Research's model-agnostic many-agent harness for long-horizon research in mathematics and theoretical computer science, by Vahab Mirrokni, Yuan Deng, Jieming Mao, and Song Zuo with David Woodruff and Honghao Lin of CMU. Rather than asking a model for a proof, Colosseum explores alternative strategies first, uses a readiness gate to decide when a route is mature enough to decompose, represents the plan as interdependent section-level subproblems, generates and attacks candidates in parallel, and routes verifier findings back to the affected part of the argument. Running on Gemini 3.1 Pro and 3.7 Flash it reaches 71% on TCS-Bench (300 research-level tasks), solves 218 of 222 Codeforces problems, and produced five results on open problems that have appeared at FOCS and in JMLR. It is the system behind Antigravity's "Teamwork" preview of 27 August 2026, whose launch example was a Lean-verified proof of Knuth's Cycles Conjecture. Filed for the harness line alongside Antigravity.

Paper

agentsagenticreasoningresearch

Related