Google's Antigravity agent teams formalize landmark math proofs and rewrite core software
Google says teams of Gemini-powered agents checked famous proofs, read chemistry data and rebuilt video and kernel code, with each result checked by other agents.
The short version
- Teamwork-LEAP produced about 1.8 million lines of Lean code to formalize the Weil conjectures, across 28,800 subagent invocations.
- Separate agents act as critics and auditors, reviewing results without seeing the original agent's reasoning.
- A 30-agent team matched molecules to NMR data with 95% top-one and 100% top-three accuracy on hard benchmarks.
- Agents rewrote the libgav1 AV1 video decoder in about 170,000 lines of Rust, ending 20% to 30% faster than the C++ original.
- Google is opening six months of free top-tier access to 10,000 academic researchers.
Why it matters
The notable shift is from agents answering a single prompt to agents doing long, checkable projects over days or weeks. The design that makes this credible is separation: the one that does the work is not the one that grades it. That is a pattern any careful reader can demand from AI claims, in math, in code or in their own work.
What Google announced
As interestingengineering.com reported, Google has widened its Antigravity system, which runs groups of specialized AI agents built on its Gemini models. Google calls the approach Teamwork. Rather than one model producing one long answer, several agents split up a problem, try different routes and test the results with computing tools.
Google points to results in three areas: formal mathematics, chemistry and software engineering. Humans stay in the loop. The work pairs researchers with the agents instead of replacing them, and the source does not say how much guidance people gave on each project.
Math a computer can check line by line
The headline result is the formalization of the Weil conjectures. These are results about how many solutions certain equations have when the numbers come from a limited, finite set. Mathematician Pierre Deligne proved them in 1974. Formalizing means rewriting the whole argument in Lean, a programming language and proof assistant, so that software can confirm every step.
Google's Teamwork-LEAP system wrote roughly 1.8 million lines of Lean and made 28,800 subagent invocations (separate calls to a helper agent) to get there. That is the size of a very large software project, produced to restate a proof that already existed.
Google also said the system reproduced a formalization of a recent proof about finite-time blowup in the three-dimensional Navier-Stokes equations. These equations describe fluid flow, and the question of whether their smooth solutions can break down in finite time has been open for about 90 years. It is tied to one of the seven Millennium Prize Problems. Note the wording: the source describes reproducing a formalization of a recent proof, not Google declaring the problem settled on its own.
A third item is the Courtade-Kumar conjecture, an information theory puzzle posed in 2013. Researchers from Google and universities used the framework to build a proof covering the general case. Gemini models helped derive mathematical bounds and produce certificates a computer can verify, and the final proof was formalized in Lean 4.
Why the agents check each other
A known weakness of AI models is that errors compound over a long chain of reasoning. Google's answer is to split roles. Several agents work in parallel, and others are assigned to challenge and verify what they produce.
Each worker agent gets only the files and tools its task needs. Critics and auditors then review the output on its own merits, without being shown how the first agent reasoned. The idea is like having an editor who reads only the finished page, not the writer's notes.
From molecules to millions of lines of Rust
In chemistry, a team of 30 specialized agents worked with a physics-informed simulator to work out molecular structures from nuclear magnetic resonance (NMR) data, a lab measurement that hints at how atoms are arranged. On difficult benchmark molecules, the right structure was the top guess 95% of the time and within the top three guesses 100% of the time.
In software, agents rewrote Google's libgav1 AV1 video decoder from C++ into about 170,000 lines of Rust, a language designed to prevent many memory-related bugs. The output matched the original exactly across a broad test suite. The first version ran about four times slower. After the agents were given profiling tools, which show where a program spends its time, and told to speed it up, the Rust version ended up 20% to 30% faster on both x86_64 and ARM processors.
A separate two-week effort rewrote the remaining C++ in the Zircon microkernel, the core of an operating system, into about 800,000 lines of memory-safe Rust. It passed kernel unit tests and booted on ARM64 hardware. Google also said an agent-built scheduling workflow cut TPU workload preemption by 24% on some hardware types, while keeping service targets and cluster utilization intact. TPUs are Google's AI chips, and preemption means a running job gets bumped.
Who can use it and what to be careful about
Google is putting Teamwork and /boost, an iterative coding workflow, into Antigravity. It has also opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier. More specialized workflows, including formal proof tools, are promised later.
Treat these as Google's own reports. The strongest claims, such as Lean-verified proofs and exact output matches in software, rest on checks that a machine can run, which is reassuring. The performance and chemistry figures come from Google's tests, and outside replication will tell more.
What to watch
- How much human guidance went into each result, and whether outside researchers can reproduce them.
- When the promised formal proof tools reach Antigravity and who gets access.
- Whether the rewritten Rust code is adopted in real products, such as Zircon.
Try it
If you try an agent on your own scripts or spreadsheets, keep the original, ask for a change list, and run old and new versions side by side on the same inputs before trusting the new one.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by interestingengineering.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
