Google has announced a major update to Teamwork, the multi-agent feature built into its AI development environment Google Antigravity. In Teamwork, multiple AI agents critique one another's proposals and keep working for hours or days. Google reports that the framework solved seven open problems in theoretical computer science, built a RISC-V CPU simulator capable of booting an operating system, and landed optimizations in the C++ library Eigen. Notably, some of these results were reproduced with the lightweight Gemini 3.7 Flash model.
Built Around a Team, Not a Single Smart AI
Teamwork was first shown at Google I/O as part of Antigravity, and users on paid plans can invoke it with the "/teamwork-preview" command. According to Google, simply running several agents side by side tends to fall apart on hard research problems: the agents agree with early mistakes and confidently build on flawed premises.
Teamwork instead makes the familiar research loop explicit: generate candidates, try to break them, and combine what survives. Humans keep control of the objectives and final acceptance, while only the iteration in between is automated.
Team structures are defined by blueprints called patterns. Five are currently available: Iterative Coding, which tackles hard-to-decompose problems through tight test-and-fix loops; Distributed Coding, which fans decomposable engineering work out to parallel workers with critic review; Long Proof, for mathematics and theoretical computer science; Self-Verification, which inserts rigorous self-checking at every step; and Document Review, for structured critique of papers and technical documents. Gemini analyzes the prompt and selects the pattern automatically.
A pattern is written as a specification rather than executable code; the framework reads it and spins up the necessary agents. The number of agents is not fixed in advance and can change during a run as the problem reveals itself.
Seven Open Problems and 71 Percent on TCSBench
The mathematics results draw the most attention. The Long Proof pattern generates many proof strategies in parallel, pairs each with a dedicated falsifier agent that hunts for flaws, and then combines the survivors in a synthesis tree. Refuted strategies remain in the process with their objections attached, on the theory that a broken route may still contain a useful idea. The chosen strategy is then decomposed into subproblems with explicit dependencies, so independent parts run in parallel while dependent parts run in order.
Google says this approach solved seven problems, including open questions posed at major venues such as FOCS and JMLR. Specifically: coreset construction for ℓp subspace approximation, a lower bound on condition number for sparse convex optimization, the complexity of similarity computation for multi-vector embeddings, theoretical guarantees for Hadamard quantization of LLMs, an independent reproduction of an existing breakthrough on the Erdős unit distance problem, a lower bound for prefix-matrix factorizations, and the first proofs for two constructions of the even case of Knuth's Cycles Conjecture.
The results were reviewed by human experts, and for Knuth's Cycles Conjecture the 40-plus-page proof was formally verified in the Lean theorem prover. Five of the seven are already published as papers on arXiv.
The results themselves were obtained with the higher-tier Gemini 3.1 Pro, but three of them were reproduced with Gemini 3.7 Flash. Google stresses that this is the first time a Flash-tier model has produced PhD-level mathematical research when given the right orchestration framework.
On benchmarks, the Long Proof pattern combining Gemini 3.7 Flash with Gemini 3.1 Pro scored 71 percent on TCSBench, an internal evaluation suite of hard problems in theoretical computer science. That beats the 67.7 percent reported in the TCSBench paper for Gemini 3.6 Flash with Gemini 3.1 Pro and is the highest score in Google's internal testing. The ability to combine Flash and Pro within a single pattern is slated for an upcoming update.
A RISC-V Simulator That Boots an OS, Built with Flash Alone
In systems work, Google describes building a cycle-level out-of-order RISC-V CPU simulator from scratch using only Gemini 3.7 Flash. The finished simulator boots the teaching operating system xv6 to a shell and runs more than 100 standard RISC-V benchmarks.
Development ran in two phases: first building the execution logic, including the out-of-order pipeline and reorder buffer, to ensure functional correctness, and then aligning cycle-level timing with a hardware-grade reference. Validation used execution results from the open-source RISC-V processor BOOM as ground truth, and the average cycle alignment error on unseen workloads came in at 0.71 percent.
A key difficulty in this kind of simulator is the "silent execution gap," a window of up to hundreds of cycles in which internal state can drift before any visible failure appears. Teamwork addressed it by running continuous lockstep co-simulation against the reference simulator Spike. To prevent cheating, the Spike source code was sandboxed so the agents could not access it directly.
Optimizations Accepted Upstream in Eigen and ParlayHash
The open-source contributions are concrete. For Eigen, a linear algebra library widely used inside and outside Google, Teamwork was given a single constraint: keep the benchmarks intact. It found an inefficient path in matrix-vector multiplication (GeMV) when the matrix has a single row or column and implemented a dedicated fast path using direct data access, SIMD instructions, and 4-way accumulator unrolling. The change went through the usual code review and was merged into upstream Eigen.
For ParlayHash, a concurrent hash table, Teamwork contributed to the ideation of "Swiss Parlay," which incorporates techniques from Swiss Table. Google reports 2x throughput on initial inserts with 64 threads, 1.5x overall throughput on a single thread, and performance close to the fastest sequential tables while using 25 percent less memory per element. This change has also landed in the upstream repository.
Summary
By switching between patterns in which multiple AI agents repeatedly critique and combine their work, Google Antigravity's Teamwork solved seven open math problems, built a RISC-V CPU simulator that boots an operating system, and landed optimizations in Eigen and ParlayHash. That several results were reproduced with the lightweight Gemini 3.7 Flash suggests the design of the collaboration matters as much as the raw capability of any single model. The Teamwork improvements will roll out to "/teamwork-preview" in Antigravity over the next few weeks.
