Google Antigravity's Teamwork Framework Solved Seven Open Problems and Scored 71% on TCSBench

Google states in its August 27, 2026 Antigravity blog post that Teamwork, a multi-agent orchestration framework, solved seven open problems in mathematics and theoretical computer science and reached 71% on TCSBench. The 71% figure comes from running Gemini 3.7 Flash together with Gemini 3.1 Pro inside the Long Proof pattern, above the 67.7% the TCSBench paper reported with Gemini 3.6 Flash and 3.1 Pro. All seven results were obtained with Gemini 3.1 Pro, and three of them were reproduced with Gemini 3.7 Flash, which Google calls the first time a Flash-tier model produced PhD-level mathematical research. Teamwork ships as /teamwork-preview in Antigravity on all paid plans and rolls out over the coming weeks.

The Framework Does Not Fix the Number of Agents in Advance

The structural core of Teamwork is that a pattern is a specification rather than an executable program. Google states that orchestration logic is decoupled from agent descriptions, that a pattern contains no orchestration code of its own, and that the framework reads the pattern and spins up the right agents automatically based on the task. Invoking /teamwork-preview has Gemini analyze the prompt and select the appropriate pattern on its own.

The more consequential detail is that agent count is not a preset. Google states that the framework dynamically decides how many agents to spawn based on task requirements, and that agent count and team structure can shift mid-run as the problem reveals itself. The post describes each campaign as a living process rather than a fixed pipeline.

Five patterns ship with the release: Iterative Coding for non-decomposable problems solved through tight agent-test-refine loops, Distributed Coding for decomposable engineering tasks fanned out across parallel workers with critic review, Long Proof for open-ended mathematics where each candidate is stress-tested before advancing, Self-Verification for depth-first mathematical reasoning with rigorous self-checking at every step, and Document Review for structured analysis and critique of papers and technical documents.

Long Proof Makes Strategies Compete Before Any Proof Gets Written

The first principle of the Long Proof pattern is that substantial effort goes into strategy selection rather than proof writing. Google states that many candidate strategies are generated in parallel, each paired with a falsifier whose sole job is to break it. Refuted routes stay in the process with their objections attached, because a broken route may still contain a useful idea.

The selected strategy is expanded into a proof plan whose subproblems carry clear goals and explicit dependencies. Independent subproblems in the resulting dependency graph run in parallel while dependent ones follow topological order. Each subproblem gets its own tournament network, where every node in the synthesis tree reads a sample of candidates together with their critiques and produces an improved solution. If a synthesized solution fails, the network reruns with the accumulated objections.

Learning also carries across rounds. Failed drafts remain available to the next attempt, verifier findings are distilled into an answer-agnostic pitfall registry, and a shared knowledge directory records proved results, useful observations, failed approaches, and relevant references. Google devotes more of its design write-up to this failure-preservation machinery than to any performance figure.

Refusing to Discard Failures Is the Real Claim Here

Google's diagnosis of multi-agent failure is the most durable claim in the August 27, 2026 post, more so than the 71% TCSBench figure. The write-up states that loosely organized agents quickly go off track, agreeing with other agents' early mistakes and building confidently on flawed ideas. Multi-agent setups can seem interchangeable at first glance, but hard research and engineering problems expose orchestration failures.

Accept that diagnosis and Teamwork's design reads as an effort to delay agreement rather than to add agents. Assigning a dedicated falsifier to every strategy, keeping refuted routes with their objections instead of deleting them, and accumulating verifier findings as an answer-agnostic pitfall list are all devices against premature convergence. This framing explains a common practitioner experience: adding parallel agents often fails to improve results because the bottleneck is filtering bad candidates, not generating more of them.

For the same reason, reading this as a model-performance announcement misses half of it. The 71% is not a Gemini 3.7 Flash score but a score produced by 3.7 Flash, 3.1 Pro, and the Long Proof pattern together, and Google separately notes that combining Flash and Pro models within Long Proof "will be available in upcoming updates." The configuration behind the headline number is therefore not one a user could reproduce on announcement day.

How to Read the Seven Results

The seven published results differ in how far they can be verified. Five have solution papers on arXiv: coresets for ℓp subspace approximation (open problem at FOCS 2025, solution arXiv:2608.26047), a conditional lower bound on condition number for sparse least-squares objectives (open problem in JMLR 2021, solution arXiv:2608.02588), maximal inner product embeddings for Chamfer similarity (solution arXiv:2607.20393), provable Hadamard quantization (solution arXiv:2608.02564), and a near-optimal lower bound for prefix-matrix factorizations (solution arXiv:2608.08238). The remaining two, the Erdős unit distance problem and Knuth's Cycles Conjecture, were published on GitHub.

Hadamard quantization is the result carrying a number: Google states the second quantization stage was eliminated, reducing the leading constant by about 5.93 times. For Knuth's Cycles Conjecture, the system produced first proofs for two simpler constructions of the even integer case, running past 40 and 70 pages respectively, with the 40-page proof formally verified in Lean. The Erdős unit distance entry is explicitly framed as an independent reproduction of an existing breakthrough, rediscovered with no internet access.

Formal verification is the sharpest dividing line in that list. A Lean-checked 40-page proof is machine-confirmed independently of human review, while the remaining results sit at arXiv preprint stage and have not been peer reviewed. Google also states that some results used higher parallelism than the default, which means reproduction cost differs from the published configuration. One formally verified proof, five preprints, and one rediscovery cannot be counted at equal weight.

The RISC-V Simulator Is Easier to Verify Than the Math

Built with Gemini 3.7 Flash, the out-of-order RISC-V CPU simulator boots the xv6 operating system to shell and successfully simulates more than 100 standard RISC-V benchmarks. Validated against BOOM hardware execution as ground truth, it achieved an average cycle alignment error of 0.71% on unseen test workloads. The work ran in two phases: agents first established microarchitectural functional correctness, including out-of-order pipelines and reorder buffers, then aligned performance characteristics against strict timing oracles.

The anti-cheating design deserves attention. Google states the Spike simulator source code was sandboxed and that Teamwork's solution maintained continuous lockstep co-simulation against the air-gapped Spike reference. The stated difficulty is what the post calls the silent execution gap, a window of up to hundreds of cycles where microarchitectural state may diverge silently before an architectural failure becomes visible.

This item is more trustworthy than the mathematics because its judge is hardware ground truth rather than a human reader. The OS either boots or it does not, cycle counts either align or they do not, and the reference implementation is isolated so answers cannot be read off. Evaluation contamination is the most common weakness in agent performance claims, and source isolation plus lockstep comparison addresses it directly. What is missing is which workload set produced 0.71% and what the maximum error in the distribution was.

The Open-Source Commits Are the Only Externally Reviewed Evidence

The Eigen and ParlayHash contributions are the results in this announcement that passed an outside maintainer. In Eigen, Teamwork identified a suboptimal GeMV implementation for matrices with a single row or column and built a dedicated fast path using direct data access and SIMD with 4-way accumulator unrolling, merged into the upstream library. The code review process ran with assistance from Gemini 3.6 Flash.

For ParlayHash, Teamwork contributed to the ideation of Swiss Parlay, incorporating optimizations from Swiss Table. The reported gains are 2x throughput on initial inserts with 64 threads, 1.5x overall throughput on a single thread, and performance close to the best sequential table, FlatHashMap, using 25% less memory per element. That improvement also landed upstream. Google states these are not benchmark-only results but real contributions accepted by external maintainers through standard open-source code review.

From an evaluation standpoint, the value of these two items lies in the gate they passed rather than the speedup they report. Self-designed benchmarks are set by the announcer, while an upstream merge requires approval from a maintainer with no stake in the announcement. In a field crowded with agent coding claims, a merged commit link is the cheapest way to bypass a reproduction argument. The scope limit is equally clear: what merged were local optimizations in two libraries, not architectural changes.

What an Engineering Team Should Check First

The first questions for an engineering team evaluating Teamwork are configuration, problem type, and cost, not the 71% TCSBench score. Configuration comes first. The seven results were obtained with Gemini 3.1 Pro, the 71% TCSBench score combines 3.7 Flash with 3.1 Pro, and Google states that combining Flash and Pro within Long Proof is coming in a future update. Nothing supports assuming that running /teamwork-preview today reproduces the configuration behind the announcement.

The second is problem type. Google states the feature is built for problems too large, too uncertain, or too open-ended for a single agent loop, and adds that basic multi-agent approaches are often enough for routine tasks. Implementing a well-specified feature is not the use case the announcement recommends.

The third is cost and time. Teamwork runs autonomously over hours or days, and some results used higher parallelism than the default. Token consumption, wall-clock time, and per-problem cost appear nowhere in the post. Seven solved open problems and the compute that produced them are two separate pieces of information, and only the first was published.

What is verified as of August 27, 2026 is that Teamwork is available as /teamwork-preview on paid plans, ships five patterns, and is evidenced by seven open problems, 71% on TCSBench, a 0.71% cycle alignment error, and upstream merges into Eigen and ParlayHash. Per-problem compute cost, run time, error distribution, and the parallelism settings needed to reproduce the results are not published.

Source: Google Antigravity blog, "Teamwork: When AI Becomes a Research Partner," August 27, 2026, by the Antigravity Team, plus the Google blog post "Pairing Google Antigravity with Gemini 3.7 Flash solves notable multi-agent math and engineering problems," August 31, 2026. Cited facts: Teamwork as a multi-agent orchestration framework first announced at Google I/O and available as /teamwork-preview on all paid plans; the five patterns Iterative Coding, Distributed Coding, Long Proof, Self-Verification, and Document Review with runtime-determined agent counts; Long Proof's competitive strategy search with falsifiers, synthesis trees, pitfall registry, and shared knowledge directory; seven open problems (ℓp subspace approximation coresets from FOCS 2025 with solution arXiv:2608.26047, sparse convex optimization from JMLR 2021 with solution arXiv:2608.02588, maximal inner product embeddings with solution arXiv:2607.20393, provable Hadamard quantization with solution arXiv:2608.02564 and a leading constant reduced by about 5.93x, the Erdős unit distance problem rediscovered with no internet access, prefix-matrix factorizations with solution arXiv:2608.08238, and Knuth's Cycles Conjecture with 40-plus and 70-plus page proofs and Lean verification of the 40-page proof); results obtained with Gemini 3.1 Pro and three reproduced with Gemini 3.7 Flash; 71% on TCSBench with 3.7 Flash plus 3.1 Pro versus 67.7% reported in the TCSBench paper with 3.6 Flash plus 3.1 Pro, with Flash-Pro combination coming in future updates; the Self-Verification pattern's lineage in the Aletheia agent and the FirstProof Challenge; the out-of-order RISC-V simulator booting xv6 to shell across 100-plus standard benchmarks at 0.71% average cycle alignment error against BOOM, with a sandboxed Spike source, lockstep co-simulation, and a silent execution gap of up to hundreds of cycles; the Eigen GeMV single row/column fast path with 4-way accumulator unrolling merged with Gemini 3.6 Flash assistance; Swiss Parlay's 2x initial-insert throughput at 64 threads, 1.5x single-thread throughput, and 25% lower per-element memory landed upstream in ParlayHash; and the note that some results used higher parallelism than the default.