@sundarpichai@GoogleAIStudio Scoring instead of generating is the right split — a routing call shouldn't pay full generation cost for a yes or no. My dispatch loop is still manual, and a 150ms scorer picking which model tier takes the next step is the piece that would actually let it run itself.
@cursor_ai Shared context is the part that matters — a new teammate walks in with the setup already done, no rebuilding from scratch. My orchestration is still manual dispatch and verify, so agents sharing plans up front and flagging conflicts mid-work is the handoff I keep doing by hand.