Google deployed ninety-three separate sub-agents to build a core operating system framework from scratch during a live demonstration at Google I/O on May 20, 2026, according to Technogic.
That demonstration showcased Antigravity 2.0, a standalone desktop application for macOS, Windows, and Linux that replaces the earlier Gemini CLI tool. Interesting Engineering reported that Google has expanded its Antigravity multi-agent system to tackle mathematical proofs, verify complex theorems, and optimize software systems using teams of specialized agents powered by Gemini models.
Software engineering tasks are moving from single-turn terminal assistants to multi-agent orchestration. Developers no longer write code alongside an asynchronous AI model in a traditional IDE workspace. Instead, they manage parallel workflows across multiple repositories simultaneously.
The system generated approximately 1.8 million lines of Lean code across 28,800 subagent invocations to formalize the Weil conjectures, originally proved by mathematician Pierre Deligne in 1974, as reported by Interesting Engineering.
According to Interesting Engineering, Google’s Teamwork-LEAP framework additionally replicated a formalization of a recent proof addressing finite-time blowup in the three-dimensional Navier–Stokes equations. That persistent mathematical question has remained open for approximately 90 years. Researchers and Gemini models also derived analytical bounds for the Courtade–Kumar conjecture, an information theory problem first proposed in 2013, generating computer-checkable certificates in the Lean 4 proof assistant language.
Independent critics and auditors examine agent outputs without relying on original reasoning chains to minimize errors. Physics-informed simulators paired with thirty specialized agents also identified molecular structures from nuclear magnetic resonance data, achieving 95% top-one accuracy on complex benchmark molecules according to Interesting Engineering.
Software engineering tasks yielded performance gains when agents were given access to profiling tools. As reported by Interesting Engineering, agents translated Google’s libgav1 AV1 video decoder from C++ into roughly 170,000 lines of Rust code, successfully reaching complete output parity across an extensive testing suite.
The initial Rust rewrite operated about four times slower than the original C++ implementation. Performance tuning by the agents ultimately made the Rust version 20% to 30% faster than the original C++ implementation on both x86_64 and ARM processors. A separate two-week campaign rewrote remaining C++ code in the Zircon microkernel into approximately 800,000 lines of memory-safe Rust, passing unit tests and booting on ARM64 hardware.
Internal infrastructure scheduling workflows reduced TPU workload preemption by 24% on specific hardware types while maintaining service-level objectives. Google is making Teamwork and the iterative coding workflow known as /boost available through Antigravity alongside free access tiers for 10,000 academic researchers.
Keep reading
