Long agent runs rot. The fix is a loop that forgets on purpose

The loop holds nothing, by design, which means every guarantee of progress is concentrated in the few seconds at the end of each iteration when the agent writes to disk ^[8]. That is a small surface to get right, and it is the only one. A worker that wakes up blank ^[4] cannot distinguish an item nobody started from an item a predecessor finished and failed to tick off.

So the thing you audit changes. With one long-lived session you watch the context. With a supervised loop you watch the fidelity of the plan. The post is candid that this is the pattern's weak point: markdown files scattered through a repo drift away from reality, and the reasoning behind a decision has no home, so it gets re-litigated every third iteration ^[11]. That is a good argument, and it is also the argument for the author's own tooling. LLMBrain, described in the same post as a hosted MCP server holding per-project docs, an issue queue and an append-only decision log ^[12], is the recommended brain, and the only production account on offer is the author's, on his own project Zenve3D ^[13]. There are no numbers here: no iteration counts, and no comparison of token spend or defect rate against a single long session.

What is testable without buying anything is the diagnosis. It is specific: retrieval degrades, instructions from fifty turns back stop binding, and the model begins treating its own earlier mistakes as ground truth, well before the window is full ^[6]. If that holds, prompt work cannot rescue it, because the prompt is sitting in the same decaying context ^[7]. That is the part of the piece that survives its own sales pitch, and it is cheap to check against your own transcripts.

The rest is bookkeeping. Starting blank means paying the orientation cost once per work item rather than once per session ^[2], because each iteration deliberately burns a fresh window on one unit of work ^[14]. That is the tax for a supervisor that never degrades, since it never remembers anything.

The correction model deserves a second read. The post concedes that individual iterations occasionally do something profoundly dumb, and leans on later iterations to fix them, comparing the whole arrangement to stochastic gradient descent for software: noisy steps, consistent direction ^[9]. Convergence there depends on the plan being the arbiter of truth and on each bad step being recoverable, which is why the surviving state is described as files on disk, commits in git and checkboxes in a plan ^[8]. Remove version control and you have a machine that repeats its mistakes with a clean context every time. The claim that people have built entire codebases overnight this way ^[10] is the part I would want a diff for.

Read narrowly, the technique is unremarkable: run the same prompt in a fresh session forever ^[3], a shape credited to Geoffrey Huntley and named for a cartoon character who is not sharp but keeps showing up ^[5]. Read as engineering, it is a statement about where correctness lives, and it moves it out of the model and into your file system.