Skip to content- Sources: primary, discussion
- Summary: Orchestrator lets Carbon run a read-only planning phase with explicit re-planning triggers, hands edits and test runs to Hydrogen, and returns control to Carbon when Hydrogen hits friction. Google says it reached near-Carbon task success on Terminal-Bench with 24% less quota and 19% faster runs, and it is rolling out to Google AI Ultra subscribers, while the Antigravity SDK can wire the two models into fixed roles.
- Why it matters: It is a vendor-built alternative to static model routers, switching models mid-trajectory instead of choosing one from the first prompt.
- Background: Static model routers guess task difficulty from the first prompt and send the whole task to one model, which breaks down when difficulty changes during a long agent run. Google released Gemini 4 Carbon and a /swarm feature for parallel agent teams earlier in the week.