Google Antigravity Solves Seven Open Problems

Google said on 31 August 2026 that Antigravity's Teamwork multi-agent framework, paired with Gemini 3.7 Flash, solved seven previously open problems in mathematics and theoretical computer science — including Knuth's Cycles Conjecture with a Lean-verified proof longer than 40 pages. The same post cites 71 percent on TCSBench, a from-scratch out-of-order RISC-V CPU simulator that boots xv6 to a shell with 0.71 percent cycle error versus hardware ground truth, and upstream performance patches in Eigen and ParlayHash.

How Teamwork orchestrates long research

Google's Antigravity Teamwork announcement describes autonomous agent teams that collaborate, critique, and iterate for hours or days on long-horizon goals. OfficeChai's secondary write-up explains patterns such as Long Proof, where candidate strategies run in parallel with dedicated critics, and Self-Verification loops that re-check a line of reasoning. Humans set the goal and sign off; the middle hours can stay unsupervised. Google positions Gemini 3.7 Flash as the first Flash-tier model to reproduce research-grade results under that orchestration, while many of the seven proofs originally came from Gemini 3.1 Pro runs.

Knuth's Cycles Conjecture is the headline. Teamwork produced multi-dozen-page proofs for simpler constructions of the even-integer case; the shorter Lean formalization was machine-checked rather than left to human eyeballing alone. Google says five of the seven results already have arXiv papers, and human experts reviewed all but the Knuth Lean proof. Other claimed wins span sparse convex optimization, provable LLM quantization, prefix-matrix factorizations, and related FOCS/JMLR-style questions — technical enough that outside mathematicians will want the papers, not just the blog scoreboard.

Systems demos that leave the whiteboard

The RISC-V simulator is Google's engineering stress test: cycle-accurate out-of-order modeling that boots the xv6 teaching OS to a working shell and runs more than 100 RISC-V benchmarks with only 0.71 percent cycle-count error against sealed hardware ground truth. Google says the reference simulator stayed sealed so agents could not copy answers. On open source, Teamwork landed Eigen SIMD fast-paths after external review and a ParlayHash Swiss Parlay design that roughly doubled 64-thread insert throughput while cutting memory per entry by about 25 percent — changes Google says are already upstream.

For developer-tooling readers, Antigravity Teamwork currently ships as a preview command on paid plans, with more patterns promised in coming weeks. That distribution detail matters more than any single conjecture: multi-agent critique loops are becoming a product surface, not only a research paper. PromptCrates' note on OpenViking as a trending context database sits in the same stack conversation — long-horizon agents need durable memory as much as clever critics.

What to verify before celebrating

Google's claims are unusually specific, which is helpful and falsifiable. Outside researchers should check the Lean artifact for Knuth, the arXiv uploads for the other six, Eigen and ParlayHash merge history, and whether the 71 percent TCSBench figure is reproducible outside Google's harness. Flash-tier reproduction of three of seven results is the efficiency story; it does not mean every open problem now falls to a cheap model overnight. Treat this as a strong vendor research report dated 31 August 2026, triangulated by OfficeChai on 1 September, not as an independent math journal acceptance.

Compared with generative-media drops such as Runway Solaris, Antigravity's pitch is formal proof and systems correctness. If the Lean proofs and upstream merges hold, Teamwork becomes evidence that multi-agent orchestration can move the research frontier — not only autocomplete the next function.

OfficeChai's walkthrough of the Long Proof pattern is useful for practitioners: agents propose strategies, critics try to break them, and failed attempts stay in memory because wrong paths can still contain reusable lemmas. That design choice is closer to research group culture than to a single chat transcript. It also explains why runtimes stretch into hours or days — the system is searching a proof space, not completing a single function.

The Erdős Unit Distance rediscovery without internet access, described in secondary coverage, is a useful honesty check: autonomous systems can reinvent recent breakthroughs when scaffolding is strong. That is impressive and cautionary. Credit for open problems should wait on Lean artifacts, referee reports, and upstream merge commits — the same evidence Google invites outsiders to inspect.

For universities and labs, the near-term question is access: Teamwork is a paid Antigravity preview, not a free public solver. Research groups should budget for Gemini quotas, human sign-off time, and independent Lean checking rather than assuming blog demos transfer one-to-one to local clusters.

Still, a Flash-tier model reproducing multiple open-problem results under multi-agent critique is a cost curve signal. If cheaper models plus better orchestration keep closing theorem gaps, the bottleneck shifts from parameter count to verification culture — Lean, referees, and upstream maintainers.

Sources

• Pairing Google Antigravity with Gemini 3.7 Flash — Google, 31 August 2026
• Google says Antigravity with Gemini 3.7 Flash solved 7 open problems — OfficeChai, 1 September 2026