Home › News
AI research · Multi-agent systems · Mathematics · November 14, 2026

Google's AI agents help verify mathematical proofs: What /teamwork really demonstrates

Google reports multi-agent progress in mathematical proofs and systems research. We explain Lean verification, headline figures, accessibility and the limits of the claims.

Editorial: AgentenCode · November 14, 2026 · Source-checked analysis
Editorial photo-composite of coordinated computing nodes and research verification workflows; not an actual Google screenshot
Original editorial illustration; not an official product screenshot or interface.

Beyond conventional chatbot answers

On November 13, Google Research described research results from Antigravity's /teamwork capability. Specialized agents examine literature, write experimental code, test hypotheses and pass work to separate checking agents. Human researchers define questions and validate their meaning.

This is a vendor's research report, not independent proof that any open mathematical problem can be solved without human input.

How the teamwork architecture works

Google describes planners, workers, critics and verification tools. Each agent receives targeted files and tools to reduce context overload across long projects. A dedicated verifier can test outputs instead of trusting the original model's explanation.

Checks range from unmodified software tests to Lean 4. A machine-checked proof is valid for its formal theorem statement; whether that statement correctly captures the informal research problem remains a human responsibility.

The headline number: 1.8 million lines of Lean

According to Google Research, agents produced a roughly 1.8-million-line formalization of the Weil conjectures over 28,800 subagent calls. These conjectures were already proven historically. The new achievement is machine-checked formalization, not the first solution of an unsolved Weil conjecture.

Google also reports more than 919,000 lines for a formalization related to Hilbert's tenth problem. Line count shows scale, not by itself proof quality, mathematical originality or peer-reviewed acceptance.

Diagram: research question, agent team, Lean checks and human review
Original AgentenCode explainer, not a vendor diagram.

Which work may be genuinely novel

The announcement separately describes results in adaptive data analysis, algorithms and graph processing. They combine model-generated exploration with human collaborators and objective checks, and must be judged claim by claim.

A useful distinction is vendor reporting, independently inspectable mathematical or experimental evidence, and subsequent scientific peer review. These are not interchangeable milestones.

Why this matters beyond mathematics

The architecture could generalize to software: one agent proposes a change, another tries to falsify it and a third runs objective benchmarks or security checks before any human approval.

A multi-agent system is not inherently more reliable. More agents can amplify shared errors and costs if the evidence gate is weak.

A realistic company pilot

Use bounded tasks with frozen datasets, explicit acceptance tests, retained logs and human sign-off. Measure successful outcomes and regressions, not token counts or the number of running subagents.

Agent security remains fundamental. Formal correctness of one artifact does not authorize access to confidential data or uncontrolled external actions.

Availability versus research demonstration

Google describes /teamwork and /boost as available in Antigravity, with additional specialized workflows still to come. The reported research campaigns were substantial collaborations, not necessarily one-click consumer features.

Access tiers, pricing, regional availability and usage limits should be checked directly in the current product rather than inferred from the research announcement.

AgentenCode verdict: Verification is the real story

What matters is making long AI workflows inspectable: intermediate artifacts, independent test suites and explicit human responsibility. That could move AI beyond plausible explanations toward traceable results.

Commercial value still depends on measured accuracy, latency, costs and oversight. This is a significant research signal, not a blanket guarantee of autonomous discovery.

Frequently asked questions

Did Google autonomously prove an unsolved conjecture?

The announcement mixes new research findings with formalizations of previously established proofs. Each claim needs separate verification.

What does Lean 4 verify?

It mechanically checks proofs of precisely formalized theorem statements; the mapping from informal claims remains important.

Is /teamwork publicly usable?

Google describes availability in Antigravity, while specialized research workflows and limits may differ.

Are more agents always better?

No. Gains require explicit tests, independent critiques, permissions and cost controls.

RESEARCH & CONTEXT

Sources for this article

This analysis relies on the original publications linked below, which readers can inspect directly.

  1. Google Research – Advancing science & engineering with Antigravity's /teamwork, 13 November 2026
  2. Google Antigravity – Teamwork research workflows and access
  3. Lean 4 – official open-source theorem prover

Checked November 14, 2026. Vendor statements, pilot scope and preview benchmarks have not been independently reproduced or certified.

NEWSLETTER · FREE · GERMAN EDITION

More signal. Less AI noise.

Our source-backed German-language digest of major AI developments, agent updates and knowledge articles. The weekly edition is being prepared.

Double opt-in: Subscription starts only after email confirmation. Your details are sent to Brevo when you submit.