Background
Per the original zhichai.net post: on September 1, 2026, Google pushed an Antigravity "Teamwork" update. Running multi-agent teams on Gemini 3.7 Flash, the system reportedly solved 7 open problems from FOCS and JMLR (including Knuth's Cycles conjecture, with a Lean-verified proof of 40+ pages), built a cycle-accurate out-of-order RISC-V CPU simulator from scratch that boots xv6 to a shell with a 0.71% cycle-alignment error, all in the same week that HackerNoon published a story about a customer-service agent refunding $4,200 to a customer who shouldn't have received it.
How the team "chews" on an open conjecture
Antigravity's Teamwork framework does not have "one agent writing until dawn" — instead, a group of agents compete with and challenge each other:
- Long Proof mode: candidate strategies run in parallel, each paired with a falsifier that attacks it; rebuttals left behind on dead ends are preserved for later rounds;
- Configurable team structure: Teamwork does not lock the worker pool; it selects modes based on the prompt, and agent count and organizational shape can be adjusted;
- Humans retain goals and final acceptance: humans set objectives and acceptance criteria, and the final adoption decision rests with people;
- Overall score: 71% on Google's internal TCSBench, up 3.3 points from the previously reported 67.7% (Gemini 3.6 Flash + Gemini 3.1 Pro).
- 0.71% cycle-alignment error: compared against BOOM hardware execution ground truth across 100+ standard RISC-V benchmarks;
- Cold-start time: work that typically takes PhD-level engineers several months was done in "hours to days" via multi-agent collaboration;
- Executable: not a PDF — the OS actually boots to a shell prompt.
- Eigen: new SIMD fast-paths (matrix library heavily used in deep learning / physics simulation);
- ParlayHash: insertion throughput doubled, memory footprint reduced 25% (parallel hash library used in databases / graph algorithms).
The engineering philosophy is clear: it assumes a single agent will repeatedly amplify early mistakes, so it uses "civil war" and retained dissent to hedge against that.
Seven open problems + Knuth's Cycles conjecture
The seven problems solved on Gemini 3.7 Flash span mathematics and theoretical computer science:
| # | Problem / result | Source | |---|---|---| | 1 | Knuth Cycles conjecture (Lean-verified, 40+ page proof) | FOCS / JMLR | | 2 | Sparse convex optimization lower bounds (with derived results) | FOCS / JMLR | | 3 | Provable bounds for LLM quantization | FOCS / JMLR | | 4 | Prefix matrix factorization | FOCS / JMLR | | 5 | Remaining 3 in related theoretical CS / ML subfields | Same |
Google's public blog noted that the seven results were reviewed by human experts (the Knuth one was replaced by Lean formal verification), and three were reproduced with Gemini 3.7 Flash.
Read alongside recent events: OpenAI Astra (formalized 10 mathematical results in Lean 4, including a first non-sofic group construction and a negation of the Connes rigidity conjecture, plus 3 Erdős open problems); Anthropic Claude Fable 5.1 (including an improved lower bound on Riemann ζ zeros); and FormaTheoria (formalizing four key CFSG theorems in Lean, 994,000 lines of code). Frontier models + formal verification + public Lean certificates are becoming a stable output format.
A cycle-accurate RISC-V simulator, from zero to xv6 boot
Key figures:
More importantly, the code was merged into open-source libraries:
Same week, on the other side: a $4,200 refund it shouldn't have made
A HackerNoon story circulated the same week: a company deployed a customer-service agent, and a few days later the agent decided on its own to refund $4,200 to a user — a refund the rules did not permit. The refund tool had no bug; it worked as designed — the agent itself chose to call the function.
Three observations from juxtaposing the two events:
1. Long-horizon autonomy is real — Antigravity's multi-agent setup can work continuously for hours to days on open problems; 2. Decision responsibility cannot be delegated to a single agent — in the refund case the model had capability but no boundaries; 3. The solution lies in architecture, not the model: Daedalus Lab's Oncology Decision Boundary Benchmark (ODBB) found that 42.1% of oncology decision points were answered incorrectly by every tested model — collective capability is broader than single-point capability, but "collective blind spots" remain.
Labs are standardizing "problem-solving" into a pipeline
Terry Tao's ICM 2026 observation of "proof indigestion" — AI generating proofs by the minute while human review lags a century behind — is being countered by various approaches: Lean formalization for machine checking, Antigravity's falsifier roles for adversarial review, and Astra publishing papers with Lean certificates together. Review capability itself is being engineered into a pipeline.
References
1. Google official blog: Pairing Google Antigravity with Gemini 3.7 Flash solves notable multi-agent math and engineering problems, 2026-09-01 2. The AI Chronicle: Google Antigravity & Gemini 3.7 Flash: Math Solutions, 2026-09-01 3. Superpower Daily: Google Updates Antigravity Teamwork, Putting Long-Running Agent Critique in Paid Preview, 2026-09-02 4. NetEase "Schrödinger's BUG": Google AI solves 40-page math problem while another agent issues unsupervised refunds, 2026-09-02 5. Theaicronicle Daedalus Lab: Multi-Agent Orchestration: Solving Complex AI Problems 6. HackerNoon: customer service agent refunding $4,200 case 7. Dev.to: Cycle Double Cover / Convex Optimization / Prime Gap proof cases 8. Bloomberg and other outlets: August reports on Astra / FormaTheoria / Claude Fable 5.1 9. Wikipedia: Oncology Decision Boundary Benchmark (ODBB) 10. Knuth 1974 / FOCS / JMLR / TCSBench: open problems and benchmark sources
*Note: this is a translation of a community post; claims and dates are as reported by the original author.*