Google AI Teams Take on Software Rewrites With Tests Guiding the Work
Google reports that coordinated AI agents rewrote its AV1 video decoder in Rust, matching output in tests and improving speed by 20 to 30 percent.
Published November 15, 2026Intelligence / Artificial Intelligence / Generative Models
A video decoder turns compressed data back into pictures. Making it faster can reduce the processing work behind a familiar daily activity. Google says a team of its AI agents has rewritten libgav1, its decoder for the AV1 video format, and improved its speed by 20 to 30 percent in testing. Google Research
The result is part of a November 13 report on Antigravity’s /teamwork, which coordinates Gemini agents on research and engineering projects. In the video experiment, Google reports matching the original decoder’s output across a broad test suite after translating it from C++ into Rust. The speed comparison covered x86-64 and ARM processors.
A rewrite must preserve years of accumulated software behavior. A development team needs the new version to handle its old responsibilities while improving something measurable: speed, memory use, reliability or maintainability. Automated checks give an AI team a way to test proposed changes against those responsibilities.
Google described Teamwork’s basic structure in August. Agents propose solutions, other agents challenge them, and the system carries useful findings into subsequent attempts. People set objectives and make the final acceptance decisions. Work can continue for hours or days, with the organization of the team changing as the problem develops. Antigravity’s August explanation
Giving each agent a defined job
The coordination system decides how to divide the assignment. Separate software components can be developed in parallel, while changes whose effects are tightly connected can pass through repeated implementation and revision. Google calls these arrangements patterns: reusable instructions defining the participating roles and the conditions for moving work forward. Teamwork’s workflow descriptions
In the decoder experiment, the first Rust version ran substantially slower than the original. Google says profiling tools then helped the agents find where it spent time and improve those parts. Profiling measures the work a program actually performs, allowing developers to concentrate on expensive operations. November report
Keeping the existing tests gives that search a stable target. For a decoder, identical output on test videos checks whether a faster implementation still produces the expected pictures. The usefulness of those results also depends on the variety of videos tested and the hardware on which the timings were collected.
The checks shape the opportunity
Mathematics offers a different kind of evaluation. Lean, an open-source proof assistant, checks whether a formal argument establishes a precisely stated proposition. Humans still have to establish that the proposition captures the problem they intended to solve. Lean
“Lean enables large-scale collaboration by allowing mathematicians to break down complex proofs into smaller, verifiable components,” mathematician Terence Tao says in an endorsement on the project’s website. That division makes a long argument easier to distribute and inspect. Lean
Software teams face a related task when they define acceptance tests. An application might produce correct answers on routine inputs while its behavior under interruptions, unusual files or scarce memory requires further examination. Choosing those cases remains an engineering judgment, and preserving their results gives later reviewers something concrete to inspect.
For engineers assessing a rewrite like Google’s decoder, the test videos, expected output and timings would provide a basis for judging each change. Those records connect the speed claim to the conditions under which the software actually worked, helping people decide when a new version is ready to use.