Google is making Antigravity’s /teamwork multi-agent workflows available to external researchers and developers. The system divides difficult projects among specialized agents, runs candidate solutions in parallel and uses independent checks such as Lean proofs, software tests, simulations, web search or hardware counters.
The design addresses two recurring weaknesses in long AI tasks. A single model attempt can be unpredictable, so /teamwork explores multiple paths and verifies the survivors. Long contexts can also dilute important details, so each agent receives only the files and tools needed for its step. Coordinators hand work to fresh agents, while separate critics review results without inheriting the original reasoning.
Google says researchers used Gemini 4 models and these workflows on mathematics, chemistry, systems software and machine-learning engineering. One collaboration addressed the Courtade–Kumar conjecture in information theory, with humans shaping the reduction and models developing analytical bounds and interval certificates. The resulting proof was formalized in Lean 4.
These are collaborative results, not evidence that agents can independently conduct reliable research. Human experts framed problems, checked assumptions and prepared results for publication. Google says /teamwork, /boost and related patterns are available in Antigravity, while additional external features will arrive in coming weeks.