Google Unveils Multi-Agent Orchestration 'Teamwork' for Antigravity

Google announced an update to the Antigravity platform featuring Teamwork, a multi-agent orchestration framework. Based on the Gemini 3.7 Flash model, this feature is designed to solve complex, long-term problems through autonomous collaboration where agents propose, critique, and refine each other's work over hours to days.

Core Structure and Patterns of Teamwork

Teamwork is not merely a collection of agents; it automatically selects and applies various Patterns based on the nature of the problem. The orchestration logic is separated from agent descriptions, allowing mechanisms validated in specific domains (e.g., adversarial critique loops) to be easily ported to other domains. Additionally, the number of agents and team structure dynamically change at runtime depending on the problem's complexity.

The key patterns currently provided are as follows:

• Iterative Coding: Solves non-decomposable problems through test-and-improve loops
• Distributed Coding: Performs decomposable engineering tasks via parallel workers and critic reviews
• Long Proof: Competitive strategy exploration for solving open-ended problems in mathematics and theoretical computer science
• Self-Verification: Deep mathematical reasoning with rigorous self-verification at each step
• Document Review: Structural analysis and critique of papers and technical documents

Achievements in Mathematics and Computer Science

In particular, the Long Proof pattern was developed for PhD-level mathematics and computer science research, ensuring that each candidate solution undergoes rigorous stress testing before advancing. This represents the porting of existing research harnesses to the Teamwork framework, demonstrating strong performance in open-ended problem solving. Teamwork is currently available via the /teamwork-preview command in all paid plans of Antigravity.