LATEST
AI Tools

Google Antigravity /teamwork: AI Teams That Prove Math, Rewrite Code

Updated November 16, 2026
14 min read
Editor's Note: This article was human-reviewed for accuracy
Complex mathematical equations and geometric diagrams written in chalk on a blackboard

Quick Answer & Key Takeaways

  • • Google's Antigravity /teamwork coordinates teams of Gemini 4 agents that propose, challenge and verify each other's work over hours or days.
  • • In results published November 13, 2026, those agent teams helped formalize a proof of the 2013 Courtade–Kumar conjecture and rewrote Google's AV1 video decoder in Rust, 20 to 30 percent faster than the original.
  • • /teamwork and /boost are available in Antigravity now, with access through Google AI Pro at $20 a month or Ultra at $100 or $200 a month.

Google's coding agents have moved past writing functions. On November 13, 2026, Google Research published results from Antigravity's /teamwork, a system that coordinates teams of specialist Gemini 4 agents against problems that stayed open for years: a 2013 conjecture in information theory, landmark theorems in pure mathematics, and production code inside Google itself. The agent teams did not just draft answers. They paired every candidate with an independent checker, a Lean 4 proof verifier, an unmodified test suite, or a hardware counter, and only the candidates that survived were kept.

The same post also opens the tools to everyone else. /teamwork and its companion /boost are available in Antigravity now, and Google has opened applications giving 10,000 academic researchers six months of free access to its highest usage tier. This article covers what the agent teams actually produced, where the headline numbers need context, and what it costs to try the system yourself.

What /teamwork actually is

Antigravity is Google's agent-first development environment, available as a desktop app, a command-line tool, and IDE extensions. /teamwork is its multi-agent orchestration layer. Instead of one agent working through a task, /teamwork organizes agents into teams with separate roles: producers draft solutions, critics look for flaws, and auditors check results from a clean context without seeing the producer's reasoning.

The design answers two known weaknesses of language models working alone. First, a single generation attempt is unreliable, so the system runs many candidates in parallel and pairs each with a dedicated falsifier, such as the Lean 4 proof checker, an unmodified software test suite, an execution simulator, or hardware counters. A candidate survives only if an independent tool confirms it. Second, models lose track of details over multi-day tasks as context grows, so /teamwork partitions context by construction: each specialized agent receives only the files and tools it needs for one step, and coordinators hand work to fresh successors while critics and auditors verify results from clean contexts.

Teamwork is not one fixed team. It expresses workflows as declarative patterns, and Gemini routes each prompt to the pattern that fits the task. The patterns available in Antigravity today include iterative coding for tightly coupled problems, distributed coding for work that fans out across parallel workers, long proof for open-ended mathematics, self-verification for depth-first mathematical reasoning, and document review for structured critique of papers. /boost packages the iterative coding pattern as a sequential solve-and-verify loop for day-to-day coding. A related feature, /swarm, launched in preview on November 10 and turns one prompt into a coordinated team of agents that plan, build, and check code in parallel on a shared message board.

The models behind the November results are Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen, with most of the flagship results built on Argon. Our coverage of Gemini 4 Argon and Gemini 4 Hydrogen has the pricing and availability detail for both.

The math results: a 2013 conjecture and the Weil conjectures

The headline result is the Courtade–Kumar conjecture, posed in 2013. It asks which Boolean function preserves the most information when a random bit string passes through a noisy channel. Researchers from Google, CUHK, USC, and Carnegie Mellon used /teamwork's Long Proof pattern to prove it in full generality. The answer is the dictator function, which just copies a single input bit. Gemini first found counterexamples to a generalized version of the conjecture, the CUHK team proposed a new candidate function that reduced the problem to a 14-variable optimization, and the agents partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates. The team then formalized the entire proof in Lean 4, producing 50 million lines of code. The proof itself appeared as a preprint in September 2026; what the November post adds is the machine-checked formalization.

The second headline is the Weil conjectures, the landmark results on counting solutions of polynomial equations over finite fields, proved by Pierre Deligne in 1974. A framework called /teamwork-LEAP, built with Google Cloud AI and paired with Gemini 4 Argon, fully autonomously produced an end-to-end, machine-checked formalization spanning about 1.8 million lines of Lean, produced across 28,800 subagent invocations. Google flags this one specifically as fully autonomous, which matters: most of the other results were produced in close collaboration with human experts.

The rest of the mathematics list is long. The system synthesized over 919,000 lines of Lean for the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers, and independently reproduced an end-to-end formalization of a recent proof of finite-time blowup for the 3D Navier–Stokes equations with smooth forcing. It generated the first formalization of the Birkhoff conjecture, open since 1927, and the first formalization of the 3D Bochner–Riesz conjecture, open for more than 50 years, finding and fixing a gap in a recently released informal proof along the way. Other items include C¹ regularity of infinity-harmonic functions in three and higher dimensions, an open question for about 20 years, and verification of informal proofs of the 40-year-old Jamison Caterpillar Conjecture.

There is also a result aimed at researchers rather than mathematicians. Ahead of STOC 2027, authors of 578 submissions, over a third of all submissions, opted in to receive Deep Review feedback on their drafts. In a survey of 70 authors, 90% rated the feedback mostly or almost all factual, 90% said they would use the tool again, and over 30% said it caught substantive proof bugs. Deep Review is implemented as /teamwork's document review pattern, and Google provides two weekly calls to it for free.

The engineering results: kernels, decoders, and a Go engine

The largest item is the Zircon microkernel. Building on an initial 25% port, a two-week Argon campaign rewrote the remaining C++ code, adding about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64 hardware. The rewrite pattern schedules each module only after its dependencies compile and links every leaf module against the original, unmodified test suite, which stops agents from stubbing out dependencies or altering tests to make a broken port pass.

Google's AV1 video decoder, libgav1, got the same treatment. /teamwork rewrote it from C++ into around 170,000 lines of Rust, achieving 100% bit-exact parity across a broad conformance test suite. The first port ran about four times slower than the original C++ code. After the team gave /teamwork access to profiling tools and a mandate to maximize speed on sample video clips, the Rust port became 20 to 30 percent faster than the original C++ and hand-optimized assembly code on both x86_64 and ARM.

Other engineering items in the post:

  • Distributed systems. Using /teamwork and Argon, Google built FarArray, a low-latency kernel-bypass system. Applications built on it scale to 1-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 implementations.
  • TPU kernels. Custom TPU kernels for graph neural networks delivered up to a 185x reduction in high-bandwidth-memory use and 8x to 15x speedups in forward and backward passes. A separate DuoDiscover pattern reached a 4.09x geometric-mean speedup across 17 JAXBench kernels and a 2.99x geometric-mean speedup across 7 hand-optimized vLLM Pallas kernels.
  • Autonomous optimization. Hill-climbing teams trained a 9x9 Go engine from scratch starting from a basic AlphaZero implementation. In under eight days the engine gained roughly 2,600 Elo, reaching 1,885 Elo, between KataGo at 16 and 32 visits per move, with a 2,300x speedup in self-play.
  • Chemistry. Thirty specialized agents in a self-correcting loop with a physics-informed spectral simulator reached 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules from NMR data, nearly doubling single-agent performance.
  • Benchmarks. Twenty-nine agents running on Argon over 11.5 hours turned 13 public tables into a 50-task causal inference benchmark, with reviewer and auditor agents vetoing hallucinated scripts.
  • Google infrastructure. Reusable /teamwork workflows identified fleet-wide memory savings surpassing Google's initial estimate of 1 PiB, and Borg's agentic scheduling workflow shipped two changes that cut TPU workload preemption by 24% on some hardware types.

What the headline numbers leave out

The results are real, and the independent checks published since November 13 are worth reading alongside Google's post, because the two biggest numbers need context.

First, the 50 million lines of Lean behind the Courtade–Kumar formalization. An independent audit cloned the public repository and counted 50,373,106 lines, so the number checks out. But the repository's own guide says only 1,359 files, about 239,000 lines, are readable proof code. The other 44,141 files are generated certificate data, numerical bounds over pieces of the parameter space that the Lean kernel checks mechanically. Under half a percent of the line count is the proof a person would read. The audit trail is still strong: a clean rebuild with a byte-identical final output, an independent run of the verification kit, and acceptance by the Lean Foundation's external checker, all between September 23 and October 15.

Second, the Weil conjectures formalization has no public code behind it yet. Google's post links to Wikipedia pages for the Weil conjectures and Pierre Deligne, with no repository, no paper, and no statement file, so the claim cannot be independently checked. A separate group, Axiom, published its own public formalization of the Weil conjectures on November 6, a week before Google's post, under an Apache 2.0 license. That does not contradict Google's claim. It means the only Weil formalization anyone can inspect today is not Google's.

Two more points of context. The Courtade–Kumar proof is not new to November: it appeared as a preprint on September 21, 2026, and independent proofs by other mathematicians landed on arXiv within a day. Google's November post mentions none of them. And the 3D Navier–Stokes item is a formalization of a recent proof about the forced equations. It is not a solution of the Millennium Prize problem, which asks about the unforced equations. None of this makes the results unimportant. A machine-checked formalization that an external checker accepts is a genuine result, and the engineering items are checked by test suites rather than by Google's word alone. The pattern is consistent: the strongest claims are the ones with public code or passing tests behind them, and the weakest are the ones where Google is the only witness.

How to try it, and what it costs

/teamwork and /boost are available in Antigravity today, alongside Deep Review. The specialized formal-proof pattern and additional workflows arrive in upcoming updates, with deep-dive tutorials promised on the Antigravity blog in the coming weeks. The feature is invoked as /teamwork-preview, and Google's August announcement says it is available on all paid plans.

There is no standalone Antigravity subscription. Access is billed through Google AI plans, and Antigravity reads the entitlement from the plan. Google's own plans page describes limits in adjectives rather than numbers, and the dollar figures below come from third-party pricing coverage of Google's May 2026 plan changes, which report US list prices. Treat them as reported, not as a Google rate card, and confirm the live price before subscribing.

PlanMonthly price (USD)Approximate GBPApproximate CADWhat it buys in Antigravity
Individual (free)$0£0C$0Rate-limited access with a weekly quota refresh
Google AI Pro$20£15.13C$28.53Higher weekly limit, quota refreshes every five hours; AI credits available as overage
Google AI Ultra$100£75.66C$142.66Five times the Pro token allowance
Google AI Ultra (top tier)$200£151.32C$285.32Twenty times the Pro token allowance

GBP and CAD figures are approximations at November 16, 2026 mid-market rates (1 USD = 0.7566 GBP, 1 USD = 1.4266 CAD). Google One lists the Ultra plans from $99.99 a month, and local pricing varies by country. The top Ultra tier was cut from $250 to $200 in the May 2026 changes. All Gemini models in Antigravity draw from one shared limit, deducted at API rates, and paid plans can enable purchased AI credits after the baseline quota runs out. There is no bring-your-own-key route for extra limits.

For API users, the model behind most of the flagship results has its own price. Gemini 4 Carbon is $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1. That matters for /teamwork in particular: a multi-agent campaign burns quota quickly, and parallel multi-agent work consumes the shared pool fast. The free tier is enough to evaluate the tool. It is not enough to run a campaign like the ones in Google's post.

Academic researchers get a separate route. Google has opened applications giving 10,000 academic researchers six months of free access to its highest usage tier. The offer requires approval rather than automatic enrollment, and it is aimed at researchers working on frontier problems rather than general use.

Five things to know before you run a campaign

Pick the problem, not the pattern. Gemini routes each prompt to a pattern automatically, and the pattern determines the team structure. A vague prompt gets a generic team. The November results all started from a precisely stated problem with a clear checker: a conjecture with a Lean statement, a decoder with a conformance suite, a kernel with a benchmark.

Give it a falsifier, not just a goal. The system works because candidates are checked by something other than the model that produced them. A proof needs Lean. A port needs the original, unmodified test suite. A kernel needs a hardware timer. Without an independent checker, a team of agents can agree with each other's early mistakes and build confidently on flawed ideas, which is exactly the failure mode Google's design is built to prevent.

Budget for the burn. Every agent in a campaign draws from the same shared token pool at API rates. A Long Proof run that spawns dozens of subagents, as the Weil formalization did with 28,800 subagent invocations, is a quota event, not a casual experiment. Start on the free tier with a small, well-specified task and watch the usage meter before committing to a multi-day run.

The formal-proof pattern is not in your hands yet. The November post says the formal-proof pattern and additional specialized workflows arrive in upcoming updates. What is available today is /teamwork and /boost with the general patterns. If your goal is Lean formalization specifically, check the Antigravity blog for the rollout before planning around it.

Verify the output the way Google did. The Courtade–Kumar formalization was accepted only after a clean rebuild, an independent verification run, and an external checker. Treat any result from an agent team the same way: rebuild it, run the tests, and have a person confirm the formal statement matches the original problem before calling it done.

FAQ

What is Google Antigravity /teamwork?

It is a multi-agent orchestration framework inside Google's Antigravity development environment. It coordinates teams of specialist Gemini 4 agents with separate roles, producers, critics, and auditors, and pairs every candidate solution with an independent checker such as the Lean 4 proof verifier or an unmodified test suite. It is available in Antigravity today as /teamwork-preview on paid plans.

Did AI really prove a math conjecture?

Partly. Researchers from Google, CUHK, USC, and Carnegie Mellon proved the 2013 Courtade–Kumar conjecture in full generality, with /teamwork's Long Proof pattern doing much of the analytical work and the Lean formalization. The proof appeared as a preprint in September 2026. The Weil conjectures were proved by Pierre Deligne in 1974; what Google's agents produced is a machine-checked formalization of that proof, and the code behind it has not been published yet.

Can I use /teamwork now?

Yes, in Antigravity, through a Google AI plan. The free Individual tier gives rate-limited access, Pro costs $20 a month, and Ultra costs $100 or $200 a month for five or twenty times the Pro allowance. The formal-proof pattern is still to come. Academic researchers can apply for six months of free access to the highest tier, with 10,000 places available.

How much does Antigravity cost?

There is no standalone Antigravity price. Access comes through Google AI plans: $0 for the free tier, $20 a month for Pro, and $100 or $200 a month for the two Ultra tiers, which carry five and twenty times the Pro token allowance. At November 16, 2026 rates, $20 is about £15.13 or C$28.53. Confirm the live price before subscribing, since Google does not publish a dollar rate card for the paid tiers.

What is Deep Review?

Deep Review is /teamwork's document review pattern, applied to academic manuscripts. Ahead of STOC 2027, authors of 578 submissions opted in to receive feedback on their drafts. In a survey of 70 authors, 90% rated the feedback mostly or almost all factual, and over 30% said it caught substantive proof bugs. Google provides two weekly calls to Deep Review for free.

The bottom line

Google's November 13 post is the strongest public demonstration yet of what multi-agent coding tools can do when every candidate has to survive an independent check. The math formalizations and the Rust rewrites are not demos. They are checked artifacts: a Lean kernel accepted one, and unmodified test suites passed the others. The honest reading is narrower than the headlines, though. The 50-million-line figure measures verification certificates more than proof, the Weil formalization cannot be inspected yet, and the formal-proof pattern that produced these results is still rolling out to external users.

For researchers and engineers with a precisely stated problem and a real checker, /teamwork is available now, and the free tier is enough to test whether the approach fits the work. Start small, give it a falsifier, and verify the output the way Google did. The tool is promising. The verification habit is what makes it useful.

Sources

14 min read
21 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more