Google has unveiled its new platform, Antigravity, featuring the core tool 'Teamwork', which allows the combination of multiple AI agents into coordinated squads to tackle tasks beyond the capabilities of a single neural network. This was reported by RBC-Ukraine, citing materials from Google Antigravity. According to the developers, each virtual assistant works independently, seeking its own solutions, which are then passed to 'critic agents' and independent auditors who verify the output from scratch without relying on the original algorithm's logic. This approach, Google claims, solves the main problem of modern LLMs — the tendency to accumulate errors during long reasoning chains.

Architecture: Task Delegation Instead of Monolith

Unlike traditional approaches where one large model attempts to deduce a solution from start to finish, 'Teamwork' strictly segments duties. Each agent receives a strictly limited set of files and tools necessary for its specific task. After a stage is complete, results are passed to subsequent agents acting as critics: they are unaware of the predecessor's reasoning and verify the output autonomously. This allows for early filtering of false reasoning paths and significantly reduces the probability of systematic errors in long chains of deduction. According to Google, this very architecture has enabled a series of results previously unattainable for single models.

Proving Theorems: From Weil Conjectures to Navier-Stokes

Among the most notable scientific achievements claimed by the Antigravity team is the formal verification of classical mathematical results. The Teamwork-LEAP system generated about 1.8 million lines of Lean code during 28,800 runs to translate the famous Weil conjectures, proven by Pierre Deligne in 1974, into a formal language understandable by computers. A separate experiment concerned the Navier-Stokes equations—one of the so-called 'Millennium Prize Problems' that the global mathematical community has been working on for about 90 years. The agents reproduced a computer verification of the proof for these equations, which, according to the developers, demonstrates the system's ability to handle tasks at the edge of computational mathematics.

Information Theory and Chemistry: 30 Agents vs. Molecular Structure

In the field of information theory, Gemini models, in collaboration with university researchers, found a complete proof for the Courtade-Kumar conjecture proposed in 2013 and formulated it in Lean 4. In a chemical experiment, a team of 30 AI agents, supplemented by a physical simulator, determined the structure of complex molecules based on nuclear magnetic resonance (NMR) data. Google states that the exact structure was guessed on the first try in 95 percent of cases, and in 100 percent of experiments, the correct variant was among the top three suggested by the system. These metrics, if confirmed by independent verification, are comparable to results previously achieved only with the participation of experienced structural chemists.

Rewriting Code: 800,000 Lines of Rust in Two Weeks

Beyond pure science, the team approach was applied in engineering. AI agents translated the libgav1 AV1 video decoder from old C++ to Rust, writing about 170,000 lines of code. Initially, the new version ran four times slower than the original, but after providing the agents with tools to identify bottlenecks, the Rust program became 20–30 percent faster than the previous one on x86_64 and ARM processors. In another experiment, the agents rewrote the remaining C++ code in the Zircon system kernel—about 800,000 lines—into Rust in just two weeks; the system subsequently passed all tests and booted on ARM64 hardware. Furthermore, AI optimization of Google’s internal servers reduced TPU interrupt latency by 24 percent.

Access for Researchers: 10,000 Scientists to Receive Six Months Free

Google is opening access to the Teamwork system and the rapid coding tool /boost via the Antigravity platform. Simultaneously, the company has announced it is accepting applications from 10,000 scientists who will receive six months of free access to the most powerful version of the service. According to the developers, this should accelerate the adoption of agentic approaches in academia and allow independent groups to verify the claimed results on their own tasks. Given that most large AI labs in 2025–2026 are focusing on individual 'super-models', Google's bet on multi-agent architecture appears to be a deliberate strategic choice aimed at solving problems where the length and complexity of the reasoning chain exceed the context window of any single model.