Product Launch3 min read

Google adds /teamwork multi-agent mode to Antigravity

Google Research has opened /teamwork in Antigravity, which runs parallel AI agents alongside automated checkers such as the Lean 4 proof checker and test suites. It is also offering 10,000 academic researchers six months of free top-tier access.

AE
Aidose Editorial
Nov 16, 2026

Google has made /teamwork, a multi-agent system in its Antigravity coding environment, available to external users, according to a November 13 Google Research blog post by Vahab Mirrokni, Google Fellow and VP of Google Research. Teamwork runs AI agents in parallel to explore candidate solutions, and dedicated verifiers check which ones work. Verifiers include the Lean 4 proof checker, test suites, simulators and hardware counters.

Each agent gets only the files and tools it needs, which Google says keeps context from sprawling over multi-day tasks. Teamwork sends each prompt to a workflow "pattern." Those include /boost, a solve-and-verify loop for coding tasks; Long Proof, for extended formal proofs; and a document-review pattern behind a feature called Deep Review. Gemini 4 models with /teamwork, /boost and Deep Review are available in Antigravity now. A formal-proof pattern and a code-rewrite pattern are described as coming soon.

Google also opened applications for 10,000 academic researchers to get six months of free access at the highest usage tier. The post names Gemini 4 Argon, Carbon and Hydrogen, and says Argon is behind most of the engineering results.

The post reports a long list of results produced with the system. These include a proof of the Courtade–Kumar conjecture in information theory, formalized in about 50 million lines of Lean 4, and an autonomous formalization of the Weil conjectures in about 1.8 million lines of Lean across 28,800 subagent calls. Google also reports rewriting the libgav1 AV1 decoder from C++ into about 170,000 lines of Rust with bit-exact output. The Rust version started about 4x slower and ended 20–30% faster than the original after profiling-driven tuning. Two shipped changes to Google's Borg scheduler cut TPU workload preemption by 24% on some hardware types.

Google notes that human experts set the goals and verified and edited the outputs, and that many results are preprints rather than peer-reviewed work. Some items formalize existing proofs rather than solving open problems. For developers, the pattern is the point: pairing parallel agents with a hard automated check, such as a proof checker, a test suite or a conformance suite, is how Google got multi-day agent runs to produce results it could trust.

AE
Aidose Editorial
Nov 16, 2026 · 3 min read
Back to News