Google Antigravity /teamwork Enables Multi-Agent Breakthroughs in Math, Science, and Systems Engineering
Answer First
Google’s Antigravity /teamwork platform, built on Gemini 4 Argon, Carbon, and Hydrogen models, has produced new proofs, algorithms, and system optimizations across mathematics, chemistry, and large‑scale engineering, and the multi‑agent tools are now publicly available for researchers and developers.
What /teamwork Is and Why It Matters
/teamwork is a multi‑agent workflow framework that pairs generative Gemini models with dedicated verification agents (Lean 4 proof checkers, test suites, simulators, etc.) to automate long‑horizon reasoning. By partitioning context, running parallel explorations, and automatically falsifying candidates, the system overcomes two major limits of current LLMs: unpredictable single‑shot generation and loss of detail over multi‑day tasks. This design enables autonomous research cycles that were previously infeasible for individual humans or single‑model pipelines.
Mathematics and Theoretical Computer Science Breakthroughs
Courtade–Kumar conjecture fully proved
- A collaboration of Google, CUHK, USC, and CMU used /teamwork’s Long Proof pattern to formalize the 2013 Courtade–Kumar conjecture in Lean 4, producing ≈50 million lines of code.
- Gemini 4 Argon identified counter‑examples to a generalized version, reduced the problem to a 14‑variable optimization, and generated analytical bounds that extended the proof from balanced to unbalanced functions without human input.
Multiple open problems solved
- Adaptive data analysis – Near‑optimal sample bounds for adaptively chosen statistical queries were proved and Lean‑verified, yielding the first space‑optimal streaming algorithm for insert‑delete data streams.
- Quantum query complexity – A new construction of n‑bit Boolean functions requiring > n/2 quantum queries settled a 2014 open question and precisely characterized overhead for symmetric functions.
- Deterministic Boolean matrix multiplication – A new algorithm answered the STOC 2024 open problem, achieving the targeted speedup over cubic time without algebraic cancellations.
- High‑dimensional Steiner spanners – Proven that constant‑factor distance approximation may need polynomially many edges, even with Steiner points, resolving a long‑standing geometric lower‑bound problem.
- Stacked sparse sketches – Broke the classical sparsity barrier for linear‑dimension matrix compression, requiring only √log r non‑zeros per row and matching a new lower bound.
- Dynamic cut sparsification – Delivered the first single‑pass streaming algorithm with near‑optimal memory, update, and reconstruction time, answering a FOCS 2022 open question.
- Online inverse linear optimization – Produced a deterministic optimal‑regret algorithm with polynomial runtime, improving on prior exponential‑time solutions.
- Integer multiplication – Generated a sub‑n log n Turing‑machine bound and an O(n log n) algorithm on a 2‑D grid whose C implementation matches GMP bit‑for‑bit up to a billion bits.
- Deque conjecture – Proved a new upper bound on splay‑tree deque operations that beats every finitely iterated inverse‑Ackermann function, with Lean 4 verification.
- Approximate leave‑one‑out CV – Demonstrated that the approximation remains accurate for non‑smooth elastic‑net penalties, extending theoretical guarantees.
Formal Verification and Manuscript Review
End‑to‑end formalizations of historic theorems
- Weil conjectures – Fully formalized in Lean 4 (≈1.8 M lines) via 28 800 sub‑agent invocations.
- Hilbert’s tenth problem (general rings) – Produced 919 k lines of Lean code for the first formal proof.
- 3D Navier–Stokes finite‑time blowup – Machine‑checked formalization of a recent breakthrough.
- Birkhoff conjecture (1927) and 3D Bochner–Riesz conjecture (≥50 yr) – First formal proofs, with the latter uncovering a gap in a published informal proof.
- Additional verified results include C¹ regularity of infinity‑harmonic functions, KdV well‑posedness, and several long‑standing conjectures in combinatorics and analysis.
Deep Review for conference submissions
- Over 578 STOC 2027 submissions (≈⅓ of all papers) received Deep Review feedback.
- In a survey (N = 70), 90 % rated feedback as mostly factual and would reuse the tool; 30 % reported catching substantive proof bugs; 14 % found technical gaps requiring > 1 hour to fix.
- Authors quoted: “Tools of this kind can help produce cleaner and more precise research, faster.”
Molecular Chemistry Discovery
- A 30‑agent pipeline using /teamwork and a physics‑informed NMR simulator achieved 95 % top‑1 exact‑match and 100 % top‑3 accuracy on challenging benchmark molecules, nearly doubling single‑agent performance.
- The workflow demonstrates how autonomous agents can accelerate structure determination, a critical step for drug discovery and protein‑folding research.
Causal‑Inference Benchmark Generation
- 29 agents running on Argon transformed 13 public tables into a 50‑task causal‑inference benchmark in 11.5 hours.
- Review and audit agents automatically filtered shortcuts, hallucinations, and wording hints, producing a high‑quality test suite for counterfactual reasoning.
Systems Engineering and Coding Achievements
Distributed systems kernels
- FarArray – A low‑latency PGAS kernel built with /teamwork and Argon, enabling lock‑free connected‑components on 1‑trillion‑edge graphs with 60‑90× throughput gains and 40‑180× speedups at 6× lower RAM.
- InstantEmbedding – Personalized PageRank implementation reduced peak DRAM by up to 14× and runtime by 11‑27×.
- Custom TPU kernels for Graph Neural Networks achieved up to 185× HBM reduction and 8‑15× speedups in forward/backward passes.
Autonomous hill‑climbing for Go
- A 9×9 Go engine trained from scratch reached ≈2 600 Elo in eight days, with a 2 300× self‑play speedup and statistical‑test‑driven champion promotion.
Rust rewrites and performance gains
- A two‑week Argon campaign rewrote the remaining C++ of the Zircon microkernel into ≈800 k lines of safe Rust, passing all unit tests and booting on arm64.
- The libgav1 AV1 decoder was ported to ≈170 k lines of Rust; after profiling‑guided optimization, the Rust version became 20‑30 % faster than the original C++/assembly implementation on both x86_64 and ARM.
/boost and DuoDiscover patterns
- /boost packages /teamwork’s iterative coding pattern; analysis of past runs showed sequential implementation outperforms parallel splits for tightly coupled code, improving token efficiency.
- DuoDiscover on Argon identified performance bottlenecks for JAX and Pallas TPU kernels, creating 23‑second proxy evaluators that achieved 4.09× geometric‑mean speedup across 17 JAXBench kernels and 2.99× across 7 hand‑optimized vLLM kernels.
Production Infrastructure Impact at Google
- Fleet‑wide memory optimization – /teamwork workflows identified opportunities exceeding the initially projected 1 PiB savings, using techniques such as struct padding elimination and scalar narrowing.
- Borg scheduling – Agentic scheduling improvements reduced TPU workload preemption by 24 % while maintaining SLOs and cluster utilization.
Availability to the Community
- Gemini 4 models, /teamwork, /boost, and Deep Review are available today in Antigravity.
- Google is offering 10 000 academic researchers six months of free access to the highest‑usage tier.
- Upcoming releases will add the formal‑proof pattern and additional specialized workflows, accompanied by tutorials on the Antigravity blog.
Community Reaction (Hacker News Comments)
"I've been finding a lot of success with team‑based workflows… role‑based, multi‑agent setups improve performance." – CharlieDigital (links to related papers).
"We already announced parts of /teamwork on the Antigravity blog; more visualizations and examples are coming soon." – mirrokni (author).
"It is not clear how /teamwork differs from Claude dynamic workflows; Antigravity feels behind and lacks an auto mode." – stbenjam.
"The post does not explain how to use /teamwork." – gdevenyi.
"The amount of breakthroughs this could deliver is staggering; a domain expert can now guide Antigravity and deliver fast results." – s_u_d_o.
These comments highlight enthusiasm for role‑based multi‑agent systems, requests for clearer documentation, and comparisons to competing platforms.
Conclusion
Google’s Antigravity /teamwork demonstrates that coordinated multi‑agent pipelines, powered by Gemini 4, can autonomously generate and verify new mathematics, accelerate scientific discovery, and optimize large‑scale systems. By open‑sourcing the harness and offering free academic access, Google is inviting the broader research community to build on these capabilities and further expand the frontier of AI‑augmented science and engineering.