Google Research has reported that multi-agent systems built on Antigravity's /teamwork feature, running on Gemini 4 models, helped researchers and engineers make progress on open problems in mathematics, chemistry, systems software and machine learning engineering. Teams from Google and academic partners, including CUHK, USC and CMU, used the Long Proof pattern to prove the Courtade–Kumar conjecture, a 2013 information theory question, in full generality. The result shows that when a random bit string passes through a noisy channel, no compression of the input to a single bit preserves more information than a "dictator function" that copies one input bit. The authors formalized the proof in Lean 4, which the post says produced about 50 million lines of code.
Other results cover optimization, statistics, quantum computing and algorithms. A multi-agent harness running on Gemini 4 Argon independently produced Lean-verified, near-optimal sample bounds for adaptively chosen statistical queries, closing a gap open since 2014. The authors say this led to the first nearly space-optimal algorithm for data streams with insertions and deletions that stays accurate against adaptive adversaries. A separate construction settled a 2014 quantum computing question. Google also reports production infrastructure and kernel optimization gains. The design rests on two ideas. First, parallel agents are checked by automated falsifiers such as the Lean 4 proof checker, unmodified test suites and execution simulators. Second, each agent receives only the files and tools it needs, which the company says limits the loss of detail over multi-day tasks. Human researchers still set the questions and constraints. Google says /boost, a standalone solve-and-verify loop, and other patterns are available in Antigravity now, with broader external features promised in the coming weeks.
For Australian businesses and professionals, this is mainly a research and developer tool announcement, and the post gives no pricing or Australia-specific availability. Costs are undisclosed, so budgeting is not yet possible. If your team already uses Antigravity or qualifies for the researcher access program, trying /boost on a bounded problem with human verification is a reasonable low-risk test. Anyone feeding client or personal data into these agents should check the vendor's data handling terms, since obligations under the Privacy Act 1988 still apply. For most businesses, nothing changes yet.
Key Points
Google says /teamwork multi-agent workflows on Gemini 4 models helped make progress on several long-standing open problems across mathematics, science and engineering.
Researchers proved the Courtade–Kumar conjecture from 2013 in full generality, with the proof formalized in Lean 4.
Verification is built in: parallel agents are checked by automated tools such as proof checkers, test suites and simulators.
Context is partitioned so each agent sees only what it needs, which Google says helps with multi-day tasks.
/boost and other patterns are available in Antigravity now, with broader external access promised in coming weeks; pricing and Australian availability were not disclosed.