An agent is processing a 3,000-token customer support ticket history before drafting a response. The most decision-critical fact (this customer is on a Premium plan, with a previously-resolved similar issue last quarter) is buried in turn 4 of an 18-turn history. The agent's drafted response treats this as a routine first-time issue.
What's the right structural fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Position matters: models attend most reliably to the start and end of long inputs, less reliably to the middle. Putting the key findings at the start of the prompt — explicitly extracted, with section headers — leverages this rather than fighting it. The detailed history is still there for reference, but the agent doesn't have to find the critical fact; the structure puts it where attention will land.
Interspersing reminders burns tokens and distracts from the actual content. Position-based attention isn't fixed by frequency — it's fixed by structure.
A larger context window means more middle for things to get buried in. Lost-in-the-middle effects scale with input length, not against it.
Already contradicted by the observation: the model missed the membership tier despite it being in context. Blaming the input writer for not making it more findable is misplacing the responsibility.