Why do agents write the plan to an external todo file instead of keeping it in context?

## The problem the file solves An agent's plan has to remain available for as long as the run lasts. On a short run, the conversation itself is fine — the plan was stated a few turns ago and is still visible. On a long run it stops being fine, for a mundane reason: the context window is finite and the run is not. As turns accumulate, older material is compacted into summaries, cleared, or simply pushed past the point where the model attends to it reliably. The plan stated at turn 3 is precisely the kind of content that gets summarized into a sentence by turn 80 — and a paraphrase of a plan is not a plan, because it loses the per-item status that made it useful. Writing the plan to a file removes the dependency on what is still in the window. The file is addressed, not remembered: the agent re-reads it when it needs it, and the cost is one small read rather than carrying the whole thing forward in every turn. ## Three jobs, one artifact **Working memory.** The file answers "what has been done, what is in progress, what is next" without the agent reconstructing it from the transcript. Reconstruction from a long transcript is exactly where agents invent completed steps or repeat finished ones. **Handoff surface.** Because the state lives outside the conversation, a new session — or a human — can pick the work up by reading the file. This is what makes long-horizon work resumable at all: the run's progress is not trapped inside one conversation's history. **Audit surface.** The file's successive versions record what the agent believed the remaining work was at each point. When something goes wrong, that history is usually more informative than the raw transcript, because it is already at the level of tasks rather than tokens. For work with any review or compliance angle — a diligence review, a change to production systems — being able to show the plan as it evolved is often a requirement rather than a nicety. ## What goes in it Keep it small and status-bearing. One line per subtask, each with an unambiguous scope, a status marker, and — where it matters — a pointer to the artifact the step produced. Pointers rather than payloads is the key discipline: the plan says `findings written to findings/auto-renewal.csv`, not the contents of that file. The plan is an index of work, not a store of results; letting tool output leak into it defeats the whole point by making it too large to re-read cheaply. The goal statement belongs at the top, verbatim. It is the single most valuable line in the file, because it is the thing an agent is most likely to lose on a long horizon and the thing everything else is judged against. ## Rewrite discipline The file only works if it tracks reality. That means updating it at completion boundaries — mark the item done, note the artifact, and only then move on — rather than writing it once at the start and treating it as a record of intent. A plan file that says three items remain when five actually do is worse than no file, because the agent trusts it: it is now confidently wrong, and it will act on that. The converse failure is over-writing: rewriting the whole file every turn burns tokens and invites the model to quietly reword items, which drifts scope. Update on completion, not on every thought. ## Limits An external plan is not a correctness mechanism. It does not make the decomposition good, it does not verify that a step marked done actually succeeded — the agent writes both the work and the checkmark — and it does not stop the agent from doing something the plan never mentioned. It buys durability and legibility, which are prerequisites for the other mechanisms rather than substitutes for them. Interviewers reward candidates who say that plainly: the file makes state inspectable, and inspectability is what everything else builds on. ## Interview framing The expected answer names the finite, lossy nature of the conversation as the driver, and then names more than one benefit — memory, handoff and audit — rather than only "the model forgets". Mentioning the pointers-not-payloads discipline and the stale-file failure mode is what distinguishes someone who has run long agent sessions from someone who has read about them. This reflects practice as of mid-2026, where externalized notes and plans are a standard part of context engineering rather than an optional flourish.