Reddit dev shares checkpoint-and-clear loop for running AI agents for days without losing state

Motiv31 · reddit · 2026-10-17

A Reddit developer detailed the full loop he uses to keep agents coherent across multi-day runs:

Checkpointing: before the context window fills, the agent writes its current state (current task, done items, next steps) to a checkpoint file; anything uncertain is recorded as "Unknown" rather than guessed. The session is then cleared and a fresh one reads the checkpoint back. The checkpoint is reference-only — live state always wins on conflict.
Fake done claims remain the weak spot: the agent writes its own done list. The author is moving toward done items that carry a verification hook (a command or URL) that the next session runs before trusting the claim.
Long-term knowledge lives separately: decisions and corrections are kept as plain markdown in git, so clearing context costs context but not knowledge.

He's asking the community how they handle the fake-done problem.

More from coding & agent