Google Research has released /teamwork, a multi-agent framework designed to navigate complex, multi-day discovery loops that have historically remained open for years. By pairing Gemini 4 models with specialized workflows, the system can solve problems across information theory, molecular chemistry, and systems engineering, effectively acting as a force multiplier for human researchers.

A graphic featuring the text Accelerating discovery with Antigravity's /teamwork
Google Research's /teamwork framework is designed to accelerate scientific discovery by coordinating multiple specialized AI agents.research.google

This image serves as the primary visual for the /teamwork framework, illustrating its role in accelerating complex discovery loops.

Overcoming Context and Reasoning Limits

Standard large language models often struggle with 'long-horizon' reasoning, where attention disperses across growing context windows during multi-day tasks. To solve this, /teamwork partitions context by construction, ensuring each specialized agent receives only the exact data required for its specific role.

The system utilizes a coordinator to hand off tasks to fresh successors. It also employs 'critics' and 'auditors' who challenge and verify results from clean contexts, independent of the original producer's reasoning, to ensure accuracy across complex chains of logic.

“The system bridges this gap by pairing parallel exploration with dedicated falsifiers (e.g., the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, or hardware counters) to automatically verify which candidate solutions work.”

— Advancing science & engineering with Antigravity’s /teamwork

Formal Verification and Mathematical Breakthroughs

By pairing Gemini 4 models with the Lean 4 proof checker, the /teamwork framework can turn informal mathematical proofs into machine-checked code. This was used to prove the 2013 Courtade–Kumar conjecture in information theory, where the AI partitioned parameter spaces and developed analytical bounds without human input.

The system also independently reproduced formalizations for the 3D Navier–Stokes equations and the Birkhoff conjecture, which had been open since 1927. These results demonstrate the system's ability to navigate vast literature and verify every logical step in a proof.

Applications in Chemistry and Engineering

Beyond pure mathematics, /teamwork supports experimental laboratory science. In chemistry, the system organized 30 specialized agents into a loop with a physics-informed spectral simulator to identify molecular structures, nearly doubling the performance of single-agent models.

In systems engineering, the framework was used to develop a deterministic algorithm for online inverse linear optimization, reducing runtime from exponential to polynomial time. It also produced a new O(n log n) algorithm for integer multiplication that matches the GNU Multiple Precision Arithmetic Library (GMP) bit-for-bit.

Human-AI Collaboration and Trade-offs

While /teamwork automates the exploration of large search spaces and the writing of custom solvers, it still relies on human experts to frame the initial questions and define the theoretical or physical constraints. The system acts as a force multiplier for researchers rather than a fully autonomous replacement for human oversight.

A notable limitation remains in the transparency of the multi-agent process; while the system can produce 50 million lines of code for a single proof, the specific computational costs and latency overhead of running dozens of parallel agents and critic loops are not yet fully detailed.