Google Updates Antigravity Teamwork, Putting Long-Running Agent Critique in Paid Preview

Google has updated Teamwork, its Antigravity framework for AI-agent teams that can work through complex tasks for hours or days. Its central bet is that difficult research and engineering work improves when agents generate competing approaches, try to break them, and refine what survives.

Teamwork organizes agents into configurable patterns for generating candidates, critiquing them, verifying results and synthesizing a solution. Google designed it for research and engineering problems where loosely coordinated agents may reinforce an early mistake; people retain control of the objective and final acceptance.

Rather than use a fixed worker pool, Teamwork selects a pattern from the user’s prompt and can alter agent counts and team structure while a job runs. The current set covers iterative and distributed coding, long proofs, self-verification and document review. It is available as /teamwork-preview on Antigravity paid plans.

71%Teamwork result

Google says Long Proof reached 71% on its internal TCSBench evaluation using Gemini 3.7 Flash with Gemini 3.1 Pro.

67.7%Earlier reported result

Google compares that result with 67.7% reported in the TCSBench paper using Gemini 3.6 Flash and Gemini 3.1 Pro.

Long Proof is the clearest example of the design. It runs proposed strategies in parallel, assigns falsifiers to attack them, and preserves objections from dead ends for later rounds. Google says it then breaks a chosen plan into dependent and independent subproblems. Google reports that the pattern addressed seven mathematics and theoretical-computer-science problems. The seven results were obtained with Gemini 3.1 Pro, while three were reproduced with Gemini 3.7 Flash. Human experts reviewed the results except the Knuth’s Cycles work, which Google says was formally verified in Lean with a proof exceeding 40 pages.

Google also tested Teamwork on systems work, where output must run correctly. It says the system built a cycle-level, out-of-order RISC-V CPU simulator that booted xv6 to a shell and simulated more than 100 standard RISC-V benchmarks. Against BOOM hardware execution ground truth on unseen workloads, Google reports an average cycle-alignment error of 0.71%.

The most tangible practical result is code accepted beyond a benchmark. Google says Teamwork-generated optimizations were merged into Eigen and ParlayHash; it reports twice the insert throughput and a 25% memory reduction for ParlayHash. The broader math and benchmark figures remain company-reported, and Google says some research runs used higher parallelism than the preview’s cost-capability balance.