Google AI agents just cracked Knuth's cycles conjecture
Google just paired its Antigravity Teamwork framework with Gemini 3.7 Flash, allowing autonomous agent teams to solve seven open math and computer science problems, including Knuth's Cycles Conjecture verified in Lean. The multi-agent setup also built an out-of-order RISC-V CPU simulator from scratch that boots xv6. Why it matters: Multi-agent systems are moving past simple coding scripts into genuine scientific discovery and complex systems engineering.
Ask a single AI model to solve a deep theoretical proof or build a hardware emulator, and it will usually hallucinate within minutes. Google's answer is giving agents coworkers and letting them run for days.
AI agents clocking in for the long haul
Google just rolled out updates to Teamwork, an orchestration framework inside Google Antigravity. Paired with Gemini 3.7 Flash, the system allows autonomous teams of AI agents to collaborate, critique, and iterate over hours or days to tackle complex, long-horizon challenges.
The results are remarkably solid. In theoretical computer science, the setup cracked seven open problems across top venues like FOCS and JMLR. Most notably, it solved Knuth's Cycles Conjecture and verified it with a 40-plus-page formal proof in Lean, alongside reaching 71% on TCSBench.
On systems engineering, the agents built a cycle-accurate, out-of-order RISC-V CPU simulator entirely from scratch. It successfully boots the xv6 operating system straight to a shell, keeping cycle alignment error down to 0.71% against hardware ground truth. The agents also landed upstream patches in core open-source libraries like Eigen and ParlayHash, doubling ParlayHash insert throughput while shaving off 25% of memory usage.
Why it matters
This looks like a compelling proof that multi-agent deliberation over long time horizons beats simply prompting a single frontier giant.
Formal proofs in Lean and cycle-accurate emulators leave zero margin for hallucination. A single wrong assumption breaks the whole run. Google's choice to use Gemini 3.7 Flash instead of an ultra-heavy flagship model is clever: by keeping inference fast and cheaper, the agents can spend days reviewing, debating, and refining each other's work.
We think this shifts the focus of agentic AI. Instead of just writing toy scripts or answering one-shot prompts, autonomous swarms that can stay on track for days are starting to handle the grueling, highly specialized work in math, compilers, and low-level systems engineering.
Sources (1)
- blog.google(opens in a new tab)Originalblog.google