Advancing science & engineering with Antigravity’s /teamwork

Google Research shared results from Antigravity's /teamwork multi-agent systems, which combine human experts with AI agents to attack open problems in mathematics, chemistry, systems software and machine-learning engineering. The work spans Google teams and academic partners, and the company says the tools are being made available externally. The post is attributed to Vahab Mirrokni, Google Fellow and VP of Google Research. The models involved are Gemini 4 Argon, Gemini 4 Carbon and Gemini 4 Hydrogen, with /teamwork sitting alongside /boost and other agents in Antigravity.
The design of /teamwork addresses two limits the post identifies in current language models. One is that a single generation may miss the correct answer, so the system pairs parallel exploration with dedicated falsifiers — the Lean 4 proof checker, unmodified software test suites, execution simulators, web search or hardware counters — to verify which candidates work. The other is that models lose detail over multi-day tasks as attention spreads across a growing context window. /teamwork instead partitions context by construction: each specialized agent receives only the files and tools it needs for a step, coordinators hand off to fresh successors, and critics and auditors check results from clean contexts without the producer's original reasoning. Workflows are stored as declarative natural-language patterns, and each user prompt is routed to the pattern best suited to it.
On the mathematics side, researchers from Google, CUHK, USC and CMU used /teamwork's Long Proof pattern to prove the 2013 Courtade–Kumar conjecture in information theory in full generality. The result holds that no compression of a noisy input bit string to a single bit preserves more information than a dictator function, which simply copies one input bit. Gemini first found counterexamples to a generalized form of the conjecture from an earlier CUHK paper; CUHK then proposed a candidate function that reduced the problem to a 14-variable optimization, plus a key analytical lower bound. Gemini 4 Argon and earlier models partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates, and extended the proof from balanced to unbalanced functions without human input. The full proof was formalized in Lean 4, yielding 50 million lines of code.
Other results in mathematics and theoretical computer science came from posing open problems directly to Gemini 4 and to harnesses including Stellar Colosseum. In adaptive data analysis, a multi-agent harness on Argon independently proved and Lean-verified near-optimal sample bounds for statistical queries chosen adaptively, closing a gap open since 2014 and giving the first nearly space-optimal streaming algorithm for insertions and deletions that stays accurate against adaptive adversaries. In quantum computing, a construction settled a 2014 question by finding the first n-bit Boolean functions needing strictly more than n/2 quantum queries at constant error. In algorithms, a deterministic result resolves the main open question from STOC 2024 on combinatorial Boolean matrix multiplication and triangle detection; a new lower bound shows that approximating all pairwise distances among n points within a constant factor can require polynomially more than n edges even with Steiner points and many hops; stacked sparse sketches break a classical sparsity barrier with only √log r nonzeros per row; and Stellar Colosseum produced the first single-pass streaming algorithm for cut sparsification with near-optimal memory, update time and reconstruction time. In optimization, a deterministic algorithm for online inverse linear optimization achieves optimal regret and cuts runtime from exponential in the dimension to polynomial. For n-bit integer multiplication, a new O(n log n) single-grid algorithm's C implementation matches GMP bit-for-bit up to a billion bits. Tarjan's 1985 Deque Conjecture saw a new upper bound beating every finitely iterated inverse-Ackermann function, verified in Lean 4, and a result shows approximate leave-one-out cross-validation stays accurate for non-smooth sparse penalties such as elastic net.
Formal verification was a major theme. The LEAP framework, developed with Google Cloud AI and paired with Argon, produced a fully autonomous end-to-end machine-checked formalization of the Weil conjectures, spanning about 1.8 million lines of Lean across 28,800 subagent invocations. It also synthesized over 919,000 lines for the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers, independently reproduced an end-to-end formalization of a recent finite-time blowup proof for 3D Navier–Stokes, produced the first formalization of the Birkhoff conjecture (open since 1927), and formalized the 3D Bochner–Riesz conjecture for the first time while finding and fixing a gap in a recent informal proof. Further machine-checked results cover C¹ regularity of infinity-harmonic functions in three and higher dimensions, Killip and Vişan's theorem on the KdV equation, and verification of informal proofs of the Jamison Caterpillar Conjecture, the Ağırseven-Ollis conjecture and an improved bound on Erdős Problem #1063.
A practical application is Deep Review. Ahead of STOC 2027, authors of 578 submissions — more than a third of all submissions — opted in for feedback on their drafts. In an ongoing survey of 70 authors, 90% rated the feedback mostly or almost all factual, 90% said they would use the tool again, over 30% said it caught substantive proof bugs or significant typos, and 14% said it found technical gaps taking more than an hour to fix. Users get two weekly Deep Review calls free, and the capability is also available as /teamwork's document review pattern.
In chemistry, 30 specialized agents organized into a self-correcting loop with /teamwork and a physics-informed spectral simulator reached 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules in determining structure from nuclear magnetic resonance data, nearly doubling single-agent performance. A similar auditing workflow built a 50-task causal inference benchmark: 29 agents on Argon worked 11.5 hours over 13 public tables, with reviewer and auditor agents catching shortcuts and hallucinated scripts and converting each into a permanent test.
On the systems side, /teamwork and Argon produced FarArray, a low-latency kernel-bypass Partitioned Global Address Space built on the SNAP microkernel, supporting lock-free Connected Components that scale to trillion-edge graphs with 60x to 90x throughput gains and 40x to 180x speedups at 6x lower RAM over TCP key-value stores. InstantEmbedding reduced peak DRAM by up to 14x and runtime by 11x to 27x over accelerated MapReduce. Custom TPU kernels for graph neural networks delivered up to 185x less high-bandwidth memory use and 8x to 15x speedups in forward and backward passes.
An autonomous hill-climbing experiment started from a basic AlphaZero implementation and trained a 9x9 Go engine from scratch. The harness ran faster self-play, achieving a 2,300x speedup, with champions promoted strictly by statistical tests on locked records; in under eight days the engine gained roughly 2,600 Elo to reach 1,885 Elo, placing it between KataGo at 16 and 32 visits per move.
Code migration work used a new /teamwork rewrite pattern, said to be coming soon to Antigravity, that schedules each module only after its dependencies compile and links every leaf module against the original unmodified test suite. Building on an initial 25% port of the Zircon microkernel, a two-week Argon campaign rewrote the remaining C++ into about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64 hardware. /teamwork also rewrote the libgav1 AV1 decoder from C++ into roughly 170,000 lines of Rust with 100% bit-exact parity; initially about 4x slower, it became 20-30% faster than the original C++ and hand-optimized assembly on x86_64 and ARM after agents were given profiling tools. For everyday coding, the /boost command packages /teamwork's iterative solve-and-verify pattern, based on analysis of past run trajectories showing tightly coupled code benefits from sequential implementation and revision rather than parallel splitting.
Why it matters: developers and product teams get access to a multi-agent harness that verifies its own output against external checkers such as Lean 4 and unmodified test suites, which is the mechanism behind many of these results. The large code-migration and kernel work suggest long-running agent campaigns can be applied to real infrastructure, though the reported speedups and formalizations still rest on human framing and review. Availability is partial: many features are in Antigravity today, with more promised for external users in coming weeks.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Education Innovation