Skip to content
Neural Daily

ResearchNews

Antigravity /teamwork: Gemini 4 agents take on open problems

🇭🇺 Olvasd el magyarul →

Google Research has collected what researchers achieved with Gemini 4 models and Antigravity's /teamwork multi-agent workflow. The list runs from a Lean 4 proof of the Courtade–Kumar conjecture to a Rust port of the Zircon kernel. /teamwork and /boost are now open to outside developers.

· 4 min read

Az Antigravity /teamwork bejelentő grafikája: Mathematics, Chemistry és Systems engineering csomópontok egy Gemini 4 /teamwork panelhez kötve
Image: Google Research · Press kit · Source

What /teamwork is

/teamwork is Antigravity's multi-agent layer, built on Gemini 4 Argon, Carbon and Hydrogen (Google Research). It targets two weaknesses. First, a single generation is unpredictable, so many candidates are produced in parallel and dedicated falsifiers filter them: the Lean 4 proof checker, unmodified test suites, simulators or hardware counters.

Second, long contexts fall apart over multi-day tasks. Each agent gets only the files and tools its step needs. Coordinators hand work to fresh successors, while critics and auditors verify from clean contexts without the producer's reasoning.

Math: Courtade–Kumar and more

The headline result is information theory's 2013 Courtade–Kumar conjecture: when a random bit string passes through a noisy channel, no one-bit compression of the input keeps more information than a "dictator function" that just copies one input bit. Researchers from Google, CUHK, USC and CMU proved it in full generality.

  • Gemini first found counterexamples to a generalized version, then the CUHK team reduced the conjecture to a 14-variable optimization problem.
  • Gemini 4 Argon and earlier Gemini models partitioned the parameter space and wrote most of the new analytical bounds and interval-arithmetic certificates.
  • The proof was extended to unbalanced functions without human input.
  • The whole proof is formalized in Lean 4, about 50 million lines (
  • [object Object]
  • ).

Fifty million lines sounds huge, but that is not the readable proof. An independent check found about 239,000 lines of actual proof code; the rest is generated numerical certificates checked by the Lean kernel (sysnote). What a reader has to trust is a 63-line statement. Two other independent proofs of the same conjecture appeared on arXiv in September, so this is not the only settlement (TSN).

The post lists more: the LEAP framework autonomously formalized the Weil conjectures in about 1.8 million lines of Lean across 28,800 subagent calls. There is progress on Tarjan's 1985 deque conjecture, and STOC 2024's main open question on combinatorial Boolean matrix multiplication is resolved. The Weil formalization is not public yet.

Systems work: Rust ports and TPU kernels

  • Zircon microkernel: a two-week Argon campaign rewrote the remaining C++, adding about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64.
  • libgav1 (AV1 decoder): about 170,000 lines of Rust with bit-exact parity. With profiling tools, the port ended up 20 to 30% faster than the original C++ and assembly.
  • TPU kernels: the DuoDiscover pattern used 23-second proxy evaluators to reach a 4.09x geometric-mean speedup across 17 JAXBench kernels.
  • 9x9 Go engine: about 2,600 Elo gained in eight days from a basic AlphaZero start.

The rewrite pattern schedules each module only after its dependencies compile, and links every leaf module against the original, unmodified test suite. That stops agents from stubbing out dependencies or editing tests to make a broken port pass.

Chemistry and Google infrastructure

For determining molecular structure from NMR data, 30 specialized agents plus a physics-informed spectral simulator reached 95% top-1 accuracy, roughly double a single agent. Inside Google, Borg scheduling cut TPU workload preemption by 24% on some hardware, and fleet-wide memory work beat its 1 PiB savings estimate. Earlier Antigravity chemistry tools: Antigravity Cheminformatics Skills.

Why it matters

/teamwork and /boost are available in Antigravity today, and Deep Review offers two free calls a week. Google is giving 10,000 academic researchers six months of free access to its top usage tier; the formal proof pattern comes later. On the underlying models: Gemini 4 Hydrogen.

For developers the lesson is the pattern, not the line count: many parallel candidates, strict automatic verifiers and narrow context per agent. Anyone building an agent pipeline can borrow that setup directly. Most of the results are not peer-reviewed yet, which is worth keeping in mind.

Key takeaways

  • The Courtade–Kumar proof is checked in Lean 4, but over 99% of the lines are generated certificates.
  • Two independent proofs of the same conjecture also exist.
  • Zircon and libgav1 Rust ports were verified against unmodified test suites.
  • /teamwork and /boost are live, and 10,000 researchers can get free access.

Related articles