Exploration paired with checks
Google Research’s November 13 update describes scientific and engineering workflows in Antigravity /teamwork. Candidate generation is paired with independent checks such as Lean 4, unmodified software tests or physical simulation. Specialized agents receive scoped context, while reviewers examine results from fresh contexts.
A chemistry example combines 30 specialized agents with nuclear magnetic resonance data and a physics-informed spectrum simulator. Google reports 95% first-candidate exact matches and 100% within three candidates on its complex-molecule benchmark. Those company-reported figures do not establish success rates for arbitrary molecules or real experiments.
Availability varies by workflow
External users can already use /teamwork and /boost in Antigravity, according to the post. Formal-proof and additional specialized patterns are still forthcoming. Applications are also open for 10,000 academic researchers to receive six months of high-quota access; eligibility is not automatic.
For developers, the transferable lesson is to define reliable tests, constraints or measurements before expanding an agent team. More exploration without dependable acceptance criteria can simply create more output to review.