Over recent weeks, teams across Google and academic partners tackled problems in mathematics, theoretical computer science, molecular structure determination, and systems engineering that had been open for years. Work combined Gemini 4 variants (Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen) with multi‑agent workflows such as Antigravity’s /teamwork (and related agents like /boost). These collaborations produced advances in information theory, formal verification, molecular discovery, kernel optimization, and several production infrastructure improvements at Google. The agentic tools are being made available in Antigravity so researchers and developers can apply them.
Approach: long‑horizon search with verification
Frontier research and systems engineering require navigating extensive literature, building custom computational pipelines, testing against physical measurements, and verifying each step. /teamwork addresses two limits of current LLMs: unpredictability of single generations and loss of detail over long tasks. It pairs parallel exploration with dedicated falsifiers (for example, the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, or hardware counters) to identify which candidate solutions actually work. It also partitions context so each specialized agent receives only the files and tools needed for a specific step; coordinators hand off to fresh successors while critics and auditors challenge results from clean contexts.
Workflows are expressed as declarative, self‑improving natural‑language patterns and prompts are routed to the pattern best suited to the task. Patterns range from distributed coding and iterative solve‑and‑verify loops (packaged as /boost) to long proof and document review workflows; many features are already available in Antigravity, with further updates planned for external users.
Mathematics and theoretical computer science results
Gemini 4 and /teamwork harnesses collaborated on large proofs and algorithm generation directly from open problem statements. Notable examples include:
-
Courtade–Kumar conjecture: Researchers from Google, The Chinese University of Hong Kong (CUHK), University of Southern California (USC), and Carnegie Mellon University (CMU) used /teamwork’s Long Proof pattern (implementing Stellar Colosseum in Antigravity) to prove the 2013 Courtade–Kumar conjecture in full generality. Gemini first found counterexamples to a generalized conjecture in CUHK’s earlier work; the CUHK team proposed a new candidate function reducing the conjecture to a 14‑variable optimization plus an analytical lower bound. Gemini 4 Argon and earlier Gemini models partitioned the parameter space, developed many analytical bounds and interval‑arithmetic certificates, and extended the proof from balanced to unbalanced functions without human input. The entire proof was formalized in Lean 4, producing about 50 million lines of code.
-
Open questions in optimization, statistics, quantum computing and algorithms: researchers posed problems directly to Gemini 4 and multi‑agent harnesses (such as Stellar Colosseum and /teamwork) and received end‑to‑end algorithms and proofs, often with Lean 4 formalizations. Examples include near‑optimal sample bounds for adaptive data analysis (yielding the first nearly space‑optimal algorithm for data streams with insertions and deletions resilient to adaptive adversaries), a quantum construction resolving a 2014 open question about query complexity, deterministic algorithms breaking a long barrier for Boolean matrix multiplication and triangle detection (achieving STOC 2024’s target speedup over cubic time), new lower bounds in high‑dimensional geometry about sparse networks, and stacked sparse sketches resolving a 2025 open question in randomized linear algebra by achieving √log r nonzeros per row.
-
Additional results span online inverse linear optimization (deterministic algorithm with optimal regret and polynomial runtime), simplifications of Turing‑machine bounds for n‑bit integer multiplication below n log n, an O(n log n) algorithm with a C implementation matching GMP bit‑for‑bit up to a billion bits, progress on Tarjan’s 1985 Deque Conjecture with a new upper bound verified in Lean 4, and proof that approximate leave‑one‑out cross‑validation remains accurate for non‑smooth sparse penalties like elastic net.
Scaling formal verification and manuscript review
Formalizing proofs in Lean 4 converts a paper’s argument into code the kernel checks line by line. The LEAP multi‑agent framework, developed with Google Cloud AI and paired with Argon, produced several machine‑checked formalizations:
- A full machine‑checked formalization of the Weil conjectures in about 1.8 million lines of Lean generated across 28,800 subagent invocations.
- A 919,000+ line Lean formalization of a recent proof of Hilbert’s tenth problem over general rings of integers.
- An end‑to‑end formalization of a recent proof of finite‑time blowup for the 3D Navier–Stokes equations with smooth forcing.
Other machine‑checked breakthroughs include the first formalization of the Birkhoff conjecture, the first formalization of the 3D Bochner–Riesz conjecture (including fixing a gap in a recent informal proof), resolution of C¹ regularity for infinity‑harmonic functions in three and higher dimensions, formalization of Killip and Vişan’s theorem on the KdV water‑wave equation in H⁻¹, and verification of various longstanding combinatorial results.
Deep Review: ahead of STOC 2027, authors of 578 submissions (over one third of submissions) opted into Deep Review. In an author survey (N = 70), 90% rated the feedback mostly or almost all factual, 90% said they would use the tool again, over 30% reported substantive proof bugs or significant typos caught, and 14% reported technical gaps that took more than an hour to fix. Deep Review is implemented as /teamwork’s document review pattern and is available in Antigravity.
Molecular chemistry discovery
Pairing specialized generator agents with independent verifier agents enabled experimental workflows. Recovering a molecule’s structure from NMR data is challenging: a single misplaced bond can distinguish an inactive look‑alike from a drug. Organizing 30 specialized agents into a self‑correcting loop with /teamwork and a physics‑informed spectral simulator achieved 95% top‑1 exact match and 100% top‑3 accuracy on complex benchmark molecules, nearly doubling single‑agent performance. This also sets groundwork for ongoing work on 3D protein and complex structure dynamics. Related science problems (including nuclear fusion and quantum algorithms) benefited from /teamwork and Argon, as described in the Gemini 4 blog post.
A similar multi‑round auditing workflow produced empirical benchmarks: starting from five human acceptance criteria, 29 agents on Argon over 11.5 hours converted 13 public tables into a 50‑task causal inference benchmark. Reviewer and auditor agents removed shortcuts and hallucinated steps, turning each task into a permanent test for counterfactual questions.
Coding and systems engineering
When paired with objective evaluators (network benchmarks, game engines, unmodified test suites, hardware timers), /teamwork campaigns can run multi‑day engineering projects.
-
Distributed systems: using /teamwork and Argon, the team built FarArray, a low‑latency kernel‑bypass Partitioned Global Address Space on the SNAP microkernel. FarArray applications such as lock‑free Connected Components scale to trillion‑edge graphs with 60x–90x throughput gains and 40x–180x speedups while using 6x less RAM than TCP key‑value stores. InstantEmbedding (Personalized PageRank) reduced peak DRAM up to 14x and runtime by 11%–27% compared to accelerated MapReduce. Custom TPU kernels for Graph Neural Networks achieved up to 185x High Bandwidth Memory reduction and 8x–15x speedups in forward/backward passes.
-
Autonomous hill‑climbing: Argon‑based hill‑climbing trained a 9x9 Go engine from an AlphaZero baseline; autonomous campaigns achieved a 2,300x self‑play speedup and roughly +2,600 Elo in under eight days, reaching about 1,885 Elo.
-
Memory‑safe Rust rewrites and /boost: a /teamwork rewrite pattern scheduled module rebuilds only after dependencies compiled and linked each leaf module against the original unmodified test suite. Extending an initial 25% Zircon microkernel port, a two‑week Argon campaign rewrote remaining C++ code into roughly 800,000 lines of safe Rust that pass kernel unit tests and boot on arm64. libgav1 was rewritten into ~170k lines of Rust with 100% bit‑exact parity on conformance tests; after profiling and optimization the Rust port became 20–30% faster than the original C++ and hand‑optimized assembly on x86_64 and ARM.
-
TPU kernel optimization: using /teamwork’s DuoDiscover pattern on Argon, four parallel agents built 23‑second proxy evaluators (yardsticks) to explore many candidates between full hardware runs. DuoDiscover achieved a 4.09x geometric‑mean speedup over reference implementations across 17 JAXBench kernels and a 2.99x geometric‑mean speedup across 7 hand‑optimized vLLM TPU‑Inference Pallas kernels.
Google infrastructure applications
Examples of production usage include:
-
Fleet‑wide memory optimization: building on initial Argon memory work, /teamwork workflows identified memory savings that exceeded an initial estimate of 1 PiB. A human‑in‑the‑loop Memory Efficiency Monitoring Agent delivered additional savings via techniques such as struct padding elimination, scalar narrowing, cliff jumping, outlining, and tagged pointers.
-
Borg agentic scheduling: /teamwork and /boost‑based agentic scheduling shipped two successful changes that reduced TPU workload preemption by 24% on some hardware types while preserving SLOs and cluster utilization.
Access and next steps
External scientists and developers can use Gemini 4 with /teamwork and /boost in Antigravity and Deep Review today. Google opened applications giving 10,000 academic researchers six months of free access to the highest usage tier. Upcoming updates will bring the formal proof pattern and additional specialized workflows into Antigravity, accompanied by deep‑dive tutorials on the Antigravity blog so that developers and researchers can generate, inspect, and iteratively refine custom multi‑agent teams. The company intends to make both the models and the tooling available more broadly.
Acknowledgements
Antigravity’s /teamwork feature reflects a Google‑wide effort with many contributors (researchers, program managers and leaders named in the project acknowledgements), as well as numerous academic collaborators who contributed to the research papers.
