The individual subagent chat windows are a really nice touch. Most agent tools just run subagents in the background and you never see what they're doing. Giving each one its own window so you can jump in and course-correct without blowing up the whole run, that's very useful. It turns it from "launch and pray" into something you can actually steer.
The tiered model selection makes a lot of sense too. How does it handle escalation though? If a task kicks off on Haiku and the first attempt doesn't land, does it bump itself up to a stronger model automatically, or does the user have to start over?
Nice work. The context efficiency numbers are legit! Going from 60K tokens to 80 is a massive difference.
Quick question: the cost attribution and audit traces are solid for tracking what happened, but what about when an agent does everything "right" and it still turns out to be the wrong call downstream? The grounding gate handles hallucinations, but if the agent picks a real tool, sends valid params, and the outcome is just... a bad decision, is there any way to unwind that? Or is that intentionally outside what the runtime is responsible for?
Also, the learning system is clever — promoting tools that work, demoting ones that don't. Have you thought about pointing that same feedback loop at the decisions themselves, not just which tool got picked? (guess that was 2 quick questions)