Google reports research gains from Gemini agent teams

What Google announced

Google Research published the results on November 13 [1]. It describes work by Gemini 4 Argon, Carbon and Hydrogen models, Antigravity agent workflows and researchers at Google and partner universities.

The headline mathematical result is a proof of the Courtade–Kumar conjecture, an information-theory problem proposed in 2013. Google says Gemini models partitioned a reduced optimization problem, developed most of the new analytical bounds and extended the argument from balanced to unbalanced functions. Human researchers supplied key reductions, constraints and final review.[1]

The associated preprint [3] describes a computer-assisted proof with computational verification records and an end-to-end Lean formalization. The preprint was posted in September, so the new event is Google’s disclosure of the broader Teamwork campaign and its external product path, rather than the first appearance of that proof.

Google also reports large formalizations of existing mathematics, molecular-structure experiments, safe-Rust rewrites and production infrastructure improvements. These are different types of evidence, so readers should assess each result against its own verification method.

Why verification matters more than agent count

The useful pattern is a loop that produces candidates, assigns other agents to attack them and sends survivors to an external check. In formal mathematics that check can be Lean. In software it can be an unchanged test suite, exact output comparison, a simulator or hardware counters.[1][2]

More agents can generate more attempts, but additional attempts only help when a test can reject persuasive mistakes. Google’s documentation assigns separate critic, challenger and auditor roles, and it gives each worker a limited file scope inside an isolated project directory.[2]

This distinction should shape adoption. A repository migration with executable tests is a better pilot than an open-ended strategy memo whose quality depends on the same model that wrote it. For subjective work, an independent human rubric remains part of the evaluation.

What builders can use now

Google’s current Teamwork documentation [2] lists the Antigravity app and CLI, a scoping interview, dedicated project directories, milestone handoffs and independent review gates. It says Teamwork is available on paid plans.

Users can invoke the general workflow for long software projects, simulations, deep research and proofs. The public system includes distributed coding, document-review and math paths, but Google says additional formal-proof and specialized rewrite workflows will arrive later.[1][2]

Deep Review [4] offers a narrower route for document checking. Google currently gives users two reviews per week and lets them supply a Gemini API key for more runs. Uploaded documents are retained for display for up to one year, according to the product page, so confidential material needs a policy check before upload.

A practical Teamwork pilot

Start with one task that has a trusted oracle and a bounded cost. The goal is to learn whether coordination improves accepted work on your repository, rather than to reproduce Google’s largest campaigns.

• Choose a task that takes one engineer at least a day and has an executable acceptance test.
• Freeze the repository revision, requirements, test suite and prohibited shortcuts before the run.
• Record model names, plan tier, elapsed time, tool calls, retries and human review time.
• Keep the existing tests outside the agents’ edit scope and add a hidden regression set when possible.
• Compare Teamwork with one capable agent on the same task, using the same acceptance criteria.
• Inspect the final diff for stubs, removed assertions, broadened permissions and unrequested files.
• Repeat on several tasks before drawing a conclusion about quality, speed or cost.

What remains uncertain

Google does not publish the total number of attempted problems, failed campaigns, tokens, compute hours or dollar cost in the announcement. Without those denominators, the report shows that the selected results were possible, but it does not establish a general success rate.[1][5]

Lean checks whether a proof follows from its formal statement and assumptions. A domain expert must still confirm that the formal statement matches the intended theorem and that the assumptions are appropriate.[3][5]

Several results are preprints or first-party reports rather than peer-reviewed findings. The strongest next evidence would be independent groups reproducing comparable outcomes with the public Teamwork product, including complete cost and failure logs.[3][5]

The public workflow may differ from the internal configuration, parallelism and model access used in Google’s campaigns. A successful internal result does not predict the outcome of a smaller paid-plan run.[1][2]

Frequently asked questions

• Did Antigravity solve the Courtade–Kumar conjecture alone? No. Google describes a collaboration in which human researchers supplied important reductions and oversight while Gemini models developed parts of the proof and certificates.[1]
• Is the proof checkable? The preprint says the full proof and numerical certificates were formalized end to end in Lean, with records available online.[3]
• Can anyone use the exact formal-proof workflow now? Google offers general Teamwork on paid plans, while its announcement says some formal-proof and specialized workflows are still coming.[1][2]
• Does a larger agent team guarantee better work? No. The reported approach depends on independent checks that reject failed candidates. Agent count without a trustworthy test can multiply errors.
• What should a software team test first? Choose a bounded migration, optimization or bug class with unchanged tests, measurable performance and a manual diff review.