知Yizhi Daily
← Back to this edition

AI

Google Research presents Antigravity /teamwork multi-agent research results

Edition 2026-11-14 · 08:00 Beijing time

Google Research describes applications of the Antigravity /teamwork multi-agent system in mathematics, chemistry, systems software, and machine-learning engineering. The system combines Gemini 4 models, parallel exploration, and specialized verifiers, enabling long-horizon agents to collaborate with human experts, and is opening related tools to external researchers and developers.

Google Research published an article on November 13 introducing Antigravity’s /teamwork multi-agent system. Google Fellow and Google Research Vice President Vahab Mirrokni said internal Google teams and academic partners used Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen, and multi-agent workflows to achieve results in information theory, formal verification, molecular structure discovery, kernel optimization, and Google’s production infrastructure.

The article says frontier research and systems engineering require reviewing extensive literature, building computational pipelines, comparing against physical measurements, and verifying every step. /teamwork lets researchers pose questions and define theoretical or physical constraints, while long-horizon agents write custom solvers, explore large search spaces, test candidate designs, and check long chains of logic.

The system is designed around two limitations of current language models. First, single generations are unstable, so the system compensates through parallel exploration and specialized “falsifiers,” such as the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, or hardware counters, to automatically verify candidates. Second, models tend to lose details in tasks spanning multiple days, so /teamwork divides context by steps, letting specialized agents receive only the files and tools they need before a coordinator hands off to a new successor agent, while critics and auditors review results from a clean context.

/teamwork expresses workflows in declarative, self-improving natural-language patterns and dynamically selects patterns according to the user’s task, including distributed coding, iterative problem-solving, long proofs, and document review. Some capabilities are already available in Antigravity, and Google says it will open more features to external users in the coming weeks.

Original sources

Google Research ↗
Back to this editionOpen interactive reader