Google Antigravity /teamwork: Open Math and Engineering Results, November 2026

On November 13, 2026, Google Research published a sweeping status post on Antigravity’s /teamwork multi-agent harness — not a single benchmark drop, but a catalog of open math, theoretical CS, formal verification, chemistry, and systems engineering results produced with Gemini 4 Argon, Carbon, and Hydrogen models plus human experts. The announcement is easy to confuse with older headlines about five Cogentic math problems; this November story is wider and more engineering-heavy.

If you build long-horizon agents, the post is a pattern library: parallel exploration, dedicated falsifiers (Lean kernels, test suites, simulators), context partitioning so subagents never inherit each other’s mistakes, and declarative workflow patterns routing prompts to Long Proof, document review, DuoDiscover, or /boost.

TL;DR: what is new vs. what we already covered

Why Google says /teamwork exists

Google names two limits of plain LLM sampling:

1. Stochastic generation — you need parallel attempts plus automatic falsifiers (Lean, tests, hardware counters, search).
2. Context drift on multi-day jobs — coordinators hand off to fresh successors; critics verify from clean contexts without the producer’s chain-of-thought.

Patterns mentioned include Long Proof, document review, DuoDiscover (fast proxy evaluators for JAX kernels), and packaged /boost iterative coding. That aligns with our August write-up of Antigravity Teamwork commands and the newer /boost deep-reasoning lane.

Google Antigravity teamwork multi-agent patterns partition context for long math and systems engineering campaigns

Mathematics and theoretical computer science (as claimed)

Courtade–Kumar conjecture

Google says a team with CUHK, USC, and CMU used the Long Proof / Stellar Colosseum pattern to prove the 2013 Courtade–Kumar conjecture in information theory. Narrative: Gemini found counterexamples to an older generalization; humans proposed a reduced 14-variable optimization problem; Gemini 4 Argon partitions parameter space and generates interval-arithmetic certificates; proof formalized in ~50 million lines of Lean 4.

Other stated TCS results

The post lists a grab bag of algorithms-from-prompt outcomes — adaptive data analysis bounds, quantum query overhead, Boolean matrix multiplication progress beyond STOC 2024, sparse sketching barriers, online inverse optimization, integer multiplication on grids matching GMP to a billion bits, progress on Tarjan’s Deque Conjecture for splay trees, and more. Each item names a conference lineage (STOC, FOCS). We did not independently verify any proof PDF or Lean artifact in this run.

For readers comparing lab math agent lore, cross-read OpenAI’s math proof volume claims and Meta Muse Spark’s six-problem post — the genre is crowded; receipts matter more than headline counts.

Lean and LEAP formalization

Google describes LEAP, built with Google Cloud AI, pairing Argon with informal-to-Lean translation. Claimed outputs include:

• Weil conjectures formalization (~1.8M lines Lean, 28,800 subagent calls)
• Hilbert’s tenth problem over rings of integers (~919k lines)
• Reproduction of a Navier–Stokes finite-time blowup formalization
• First formalization of Birkhoff conjecture, 3D Bochner–Riesz, and other harmonic-analysis targets

This is not the same product surface as Gemini in Google Classroom protests. It is research infrastructure aimed at shrinking peer-review load — a theme Google continues in STOC 2027 Deep Review, where it says 578 submissions opted in and early surveys report proof bugs caught.

Science and benchmark factory

Molecular NMR structure ID: Google claims 95% top-1 exact match on complex benchmarks using 30 specialized agents in a verifier loop with a physics-informed simulator.

Causal benchmark construction: 29 agents over 11.5 hours turned 13 public tables into a 50-task causal identification benchmark with auditor vetoes — an example of using agents to build evaluation data, not just solve it.

Systems engineering results

The engineering half may matter more to daily builders than lemma counts:

• FarArray on a SNAP microkernel — trillion-edge graph components with reported 60×–90× throughput gains vs TCP stores (details in a linked Antigravity post within Google's blog chain).
• Go engine hill-climb from AlphaZero baseline to ~1,885 Elo in under eight days with statistical champion promotion.
• Zircon microkernel Rust rewrite — ~800k lines, passing kernel tests and arm64 boot, using a pattern that schedules modules only after dependencies compile against unmodified test suites.
• libgav1 → Rust with profiling-guided speedups (20–30% faster than C++ claim after optimization).
• DuoDiscover proxy evaluators cutting JAX kernel search from 7–32 minutes to 23 seconds, validated against full hardware runs.

Those stories rhyme with Tsinghua’s algorithm breakthrough coverage and NVIDIA Snapshot serving work: verify fast, promote carefully.

Long-horizon agent workflows use critics and formal checkers to falsify candidate math and code solutions

What builders can use now

Google states:

• Gemini 4 with /teamwork and /boost in Antigravity and Deep Review today.
• Two free Deep Review calls per week for authors (per the post).
• 10,000 academic slots for six months of highest usage tier.
• Tutorials forthcoming on Antigravity blog for custom multi-agent teams.

Practical takeaway: if you already run Antigravity CLI subagents, treat /teamwork as orchestration grammar — pick a pattern (Long Proof vs /boost) before you pick a model. Mixing them blindly burns tokens.

For Gemini 4 launch context see Hydrogen comparison and Hydrogen use cases.

Pattern catalog for harness designers

Google’s post names explicit patterns. Mapping them to builder mental models:

Long Proof / Stellar Colosseum — partition a giant proof into substeps with Lean or interval arithmetic as falsifier. Use when the verifier is sound but slow. Related to Claude Gödel notebooks culturally, but Google’s scale claims are orders of magnitude larger.

Document review — parallel rubric judges on manuscript segments. Already productized as Deep Review; the STOC 2027 survey stats (90% factual, 30%+ caught substantive bugs) are vendor-reported but directionally plausible for overloaded peer review.

DuoDiscover — cheap proxy evaluators validated against expensive hardware truth. This is how you should run agent kernel tuning without melting budgets — same philosophy as running fast-proxy inference stacks before full hardware truth paths.

Rewrite pattern — schedule modules only when dependencies compile; never weaken tests to greenwash. Essential lesson for anyone migrating C++ to Rust under agent assistance.

If you only adopt one idea, adopt critic-in-clean-context: never let the generator grade itself without a separated thread.

STOC Deep Review and peer-review pressure

Google says 578 STOC 2027 submissions opted into Deep Review feedback, with 70 survey respondents so far. Even if the percentages move, the mechanism matters: conferences are using agents as lint for proofs, not as authors. That shifts reviewer labor from “find the missing lemma” to “confirm the agent didn’t hallucinate a reference.”

Watch for dual-use: the same tooling that catches bugs also homogenizes style, which can obscure novel notation. Teams publishing through Antigravity should still ask human reviewers to sign off on theorem statements, not just compile logs.

Academic access and Antigravity roadmap

Google advertises 10,000 researchers receiving six months of top-tier Antigravity usage. For PhD labs comparing against Simons Foundation math initiatives, the bottleneck remains artifact release: Lean files, datasets, and error logs matter more than press claims.

Upcoming Antigravity blog tutorials on custom teams are the practical on-ramp. Until those land, read existing command docs: /plan, /boost, and the managed-agents filesystem surface.

Competitive landscape this month

Readers should cross-check any headline count against public artifacts — see open math rumor checklists.

What to believe — and what to check

Google's post is primary source marketing tied to real collaborators and named benchmarks, but it is not a neutral audit. Before you cite a result:

1. Look for public Lean artifacts, DOI papers, or repo tags — not slide claims alone.
2. Separate human-Gemini collaborations from fully autonomous lines; the Courtade–Kumar story is explicitly collaborative.
3. Compare against how to read AI benchmark hype when STOC/FOCS names appear without links.

Instrument your own /teamwork runs with the same discipline: save transcripts, diff Lean or test outputs, and tag which subagent produced each artifact so you can audit failures without replaying entire sessions.

Finally, remember that human-AI collaboration is the stated default in the November post — even “autonomous” Lean runs sat inside Google workflows with expert sign-off on theorem statements. Classroom deployments demanding the opposite lesson should read Gemini on Chromebook protests for how hype outruns governance when claims hit parents before peer review finishes.

Related reading

• Antigravity Teamwork framework (August)
• Gemini Cogentic five problems
• Antigravity /boost command
• Deep Review for papers
• Gemini 4 Hydrogen vs rivals
• Multi-agent orchestration patterns
• Google source: Research blog, Nov 13, 2026

Claims reflect Google's November 13, 2026 post unless independently verified.