The contract you pasted is now giving orders

The usable part of an LLM’s context window stops growing long before the advertised number does. A developer writing about this borrows a framing from a video: split the window into two zones, a smart zone where the model is sharp and a dumb zone where attention drops off and the model starts forgetting what you told it a few minutes ago. The handover sits somewhere around 100,000 tokens, and it does not move when the vendor prints a bigger number on the box.

That number keeps getting bigger: 200k, then 1M, then 2M. The architectures behind those windows work, in the sense that the model will accept the tokens and return output. What they do not do is keep attention evenly sharp across the whole range. The RULER benchmark and Chroma’s write-up on context rot land on the same point: effective context is a fraction of the advertised window, and performance degrades gradually as you fill it rather than dropping off a cliff at some clean limit.

For anyone running a coding agent, the effect is concrete. A modern agent spends tokens fast. A few file reads, one long debug loop, a verbose test run, and the session is past 100k before lunch. The advertised million-token window implies there is room to keep going; the part of it that is actually working for you ran out a while ago.

Agents have started to handle this themselves. Claude Code and similar tools auto-compact: when a session gets long, the agent summarizes the history so far and starts the next stretch from that summary. It helps, and it has two weaknesses worth knowing. Compaction fires after the session has already spent time in the degraded zone, so the summary is drawn from a window that was already losing detail. And the summary is produced by the same model working in that same degraded state, which gives you a lossy compression written by the component you trust least at that moment.

The developer’s own habit is to not let a session get there. When a piece of work is done, they open a fresh session and hand it a spec they wrote themselves. A spec you write is a higher-signal handoff than any automatic summary, because you decide what carries forward instead of letting a degraded model guess. It is the breadcrumb idea applied to agents: leave an artifact clean enough that the next session, or the next person, can pick it up without the history.

This scales into a workflow. Projects like obra/superpowers and mattpocock/skills build entire agent setups around small, named artifacts: PRDs, plans, skills, sub-agent handoffs. Each one moves information out of the live session and into something a later session can read on demand. The session stays small and inside the smart zone, and the durable material lives in files.

The practical version is to treat the context window as a budget rather than a container. Assume only the first chunk is really working for you. Everything you can push out of the live session into a written artifact is one less thing competing for the model’s attention, and one less thing that quietly rots as the token count climbs.