Infrastructure for AI sessions. One API for transcript, model, and compute. Run any agent on any compute — and switch mid-session without losing the thread.