Long interactions degrade in three distinct ways, and the exam expects you to name which one a scenario is describing:
- Progressive summarization loss — summarizing condenses exactly the wrong things. Numerical values, percentages, dates, and customer-stated expectations are exactly the kind of specific detail that summarization smooths away in favor of gist.
- Lost-in-the-middle — models reliably process the beginning and end of long inputs, and may omit or under-weight sections in the middle. A fact stated once, mid-transcript, is at real risk of being dropped from later reasoning.
- Tool-result accumulation — verbose tool outputs consume tokens disproportionate to their relevance. A single lookup returning 40+ fields, when only 5 matter, compounds badly across a long session.
The highest-yield remedy: a persistent case-facts block. Extract transactional facts — amounts, dates, order numbers, statuses — into a dedicated block that's included in every prompt, sitting outside the summarized history rather than inside it. This is the single most testable pattern in this section, because it directly answers the discriminator below.
Other remedies, situational rather than universal:
- A separate, structured context layer for issue data in multi-issue sessions, rather than letting each issue's facts live only in prose history.
- Trimming verbose tool outputs to relevant fields before they accumulate, not after.
- Position-aware ordering — key findings stated first, supporting detail placed under clearly labeled headers — to work with lost-in-the-middle rather than against it.
- Requiring subagents to include metadata in structured outputs, so downstream consumers don't need to re-derive it from prose.
- Returning structured data instead of verbose reasoning/content when a downstream agent has a limited context budget.
The discriminator: facts that must survive cannot live inside the thing that compresses them. If a scenario's failure is "the agent forgot the customer's original ship date," the fix is never "summarize better" — it's moving that fact somewhere summarization doesn't reach.
Traps to recognize:
- "Summarize more often" — this loses values faster, not slower, since each summarization pass is another lossy compression step.
- "Pass the full transcript" or "use a larger-context model" — both defer the problem to a bigger container rather than solving it; tool-result accumulation and lost-in-the-middle both still occur inside a large window, just later.