Context is the discipline of managing what the model can see: supplying the right files and facts, clearing or compacting when a session drifts, using memory across sessions, and recognizing when the window itself — stale plans, dead ends, bloated logs — has become the bug.
Abandoned approaches that keep getting referenced, giant logs pasted wholesale, contradictory instructions layered over hours, and stale “facts” from earlier in the session. The model weights what it sees; if what it sees is noise, you get confident noise back.
The tell is repetition: the agent re-suggests an abandoned approach, re-reads the same files, or contradicts a decision already made. At that point more prompting adds contradiction, not clarity. Clear, restate the current goal plus the few facts that survived, and continue.
Memory is for what stays true across sessions: conventions, decisions, gotchas, who is who. The window is for what is true right now. Writing session-specific state into memory is how agents confidently act on last month’s facts.
Tokens are the meter. Long unfocused sessions re-send the same bloated context every turn; focused sessions with warm caches serve most input at a fraction of the price. The cheapest token is the one you did not need to send twice.
Hours into a session, the agent keeps proposing an approach you abandoned this morning. Best move?
C. C removes the poison instead of arguing with it. A and B add more of what caused the drift; D pays more to process the same contradictions.