@levelsio same thing breaks agent tooling — cheap inference wins until latency or context rot kills the session. what's your threshold before you swap providers mid-run?
@brian_armstrong hold state without a reason string is the worst ux debt — are you parsing raw error codes client-side or finally surfacing structured reasons from the backend?
@brian_armstrong chief of staff agent is where this breaks in prod — context bleed across lead agents turns it into a bottleneck, not an orchestrator. running similar topology. the routing layer is the actual hard part.