Google opens Antigravity's /teamwork agents, claims math and code results
Google Research says its multi-agent system helped prove a 2013 conjecture and port large codebases to Rust, and offers 10,000 academics free access. Results are Google's own.
Key takeaways
- Google made its /teamwork multi-agent workflows in Antigravity available to outside scientists and developers.
- Google is taking applications for 10,000 places giving academic researchers six months free on Antigravity's highest usage tier.
- Google says /teamwork helped prove the Courtade-Kumar conjecture from 2013, with a Lean 4 formalization.
- Google says agents rewrote the remaining C++ of its Zircon microkernel into about 800,000 lines of safe Rust in two weeks, after an earlier partial port.
- All results are self-reported; independent replication and usage costs are not yet available.
Google Research has made its /teamwork multi-agent workflows available to outside scientists and developers in Antigravity, its agentic development app, and published a list of results it attributes to the system, in a Nov. 13 post by Google Fellow Vahab Mirrokni. Academic researchers can also apply for one of 10,000 places offering six months of free access to Antigravity's highest usage tier, according to the post and Data Studios.
Every result in the post is reported by Google and its academic partners themselves. Data Studios notes that most come without an independent baseline, and Another Coding Blog's newsletter lists, as a point to watch, whether the main math results hold up once outside mathematicians go through the formal proofs.
What is /teamwork?
Google says /teamwork runs teams of AI agents on Gemini models, now including Gemini 4, on long, multi-step tasks. According to Google, it pairs many parallel attempts with dedicated checkers, such as Lean 4's proof checker, original test suites and simulators, which test which candidate solutions actually work. To stop agents losing track over multi-day tasks, it gives each agent only the files and tools it needs for one step, and separate critic agents review results from a clean context, Google says. Google packages a sequential solve-and-verify loop as /boost.
What results does Google claim?
The post's main mathematics claim is a full proof of the Courtade-Kumar conjecture, an information theory problem posed in 2013, by researchers from Google, CUHK, USC and CMU, according to Google's post. Google says Gemini found counterexamples to a broader version of the conjecture, the CUHK team proposed a new approach that reduced it to an optimization problem, and Gemini models then developed most of the new analytical bounds. Google says the models also carried the proof over from balanced to unbalanced functions on their own, and the CUHK team supplied a key lower bound. The full proof was formalized in Lean 4, producing about 50 million lines of code, Google says.
Google also says its formal-verification pipeline, LEAP, produced a fully autonomous, machine-checked Lean formalization of the Weil conjectures, landmark results about how many solutions polynomial equations have over finite fields, whose proof Pierre Deligne completed in 1974. Google puts that work at about 1.8 million lines of Lean, generated over 28,800 subagent calls, and Data Studios and Another Coding Blog repeat Google's figures.
In software, Google says that, building on an earlier partial port, a two-week campaign rewrote the remaining C++ code of its Zircon microkernel into about 800,000 lines of safe Rust, which it says passes every kernel unit test and boots on Arm hardware. Google also says a Rust rewrite of its AV1 video decoder, libgav1, started out about four times slower than the original but ended up 20 to 30% faster than the original C++ and hand-tuned assembly after the agents got profiling tools and were told to optimize speed on sample video clips. Google also reports gains inside its own infrastructure, including its Borg scheduler.
What are the caveats?
According to Data Studios, the algorithm and systems results lack independent replication and several have no baseline to compare against. Data Studios also says a reported Go engine rating rests on one game, while Google's post says only one showcase game was played for the post and that the rating came from statistical tests. Data Studios also says Google has not disclosed how much usage a /teamwork run takes, how researchers will be selected or by when, or when the code rewrite and formal proof workflows behind some of the biggest claims will reach users. Google notes that a compiling Lean proof settles the original problem only once human mathematicians confirm that the formal statement matches it.
Who can use it?
Google says outside researchers and developers can already run /teamwork and /boost on Gemini 4 in Antigravity, along with its Deep Review document-review tool, and that the formal proof pattern and other workflows will follow in the coming weeks. Data Studios reports that /teamwork and /boost run on paid Antigravity plans, while Deep Review offers two free reviews a week.
Questions readers are asking
What is Antigravity's /teamwork?
A feature in Google's Antigravity app that runs teams of Gemini-based agents on long tasks, pairing parallel attempts with automatic checkers such as the Lean 4 proof checker and test suites.
Have the results been independently verified?
Not yet. The results come from Google and its academic partners. Lean proofs are machine-checked, but outside mathematicians still need to confirm the formal statements match the original problems.
Who can apply for free access?
Academic researchers can apply; Google is offering 10,000 places with six months on Antigravity's highest usage tier at no cost.
Is /teamwork free?
According to Data Studios, /teamwork and /boost run on paid Antigravity plans; Deep Review offers two free reviews a week.
Sources
- Google Research: Advancing science & engineering with Antigravity's /teamwork
- Data Studios: Google Antigravity: 6 months free top-tier access for 10,000 researchers, /teamwork
- Another Coding Blog: Another Daily AI Newsletter - November 14
- LAVX News: Google Research details Antigravity /teamwork results across math, science, and engineering
- #antigravity
- #multi-agent
- #formal-verification
- #gemini
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Agents3 min
OpenAI launches ChatGPT Wallet so its dots agents can buy for users
The payment feature starts with US ChatGPT Pro subscribers. Users approve each purchase with a biometric check, and agents get a payment token, not the real card number.
AnalysisAgents3 min
How big shopping and travel sites now write rules for AI agents
eBay, Target, Booking, Delta and Shopify have all added terms about AI agents. NextFish compared their terms with archived copies to see what changed and when.
Cybersecurity3 min
Google Project Zero cuts bug disclosure deadline to 30 days by 2028
The Google team says AI has sped up attacks. It drops the extra 30-day patch soak time now and will cut vendors' deadline from 90 to 60, then 30 days.
Big Tech2 min
Google lets Gemini read Google Wallet data to answer money questions
Google is rolling out a Wallet connection for US Gemini users, and optionally bank and card accounts via Plaid, for budgeting help. Google says access is read-only and opt-in.