A multi-agent claims assistant in Microsoft Foundry keeps a rolling conversation context using a fixed-size sliding window. Over long claims investigations, the policy-number entity and claimant name that were established in the first few turns disappear from later agent prompts, and downstream agents begin inventing identifiers. Offline evaluation shows retrieval still finds the correct claim, but the writing agent no longer mentions the original entities. Which diagnosis best explains the failure mode, and what should the team address first?
Select an answer to reveal the explanation.
Short Explanation
The correct answer is A. This is classic sliding-window amnesia: a hard cap on how many recent turns stay in the prompt quietly drops the early turns where the policy number and claimant name were first established, so later agents never see those entities again and start making them up. Retrieval can still work because the index is separate from the chat window, which is why offline RAG looks fine while the writing agent drifts. Fix entity continuity first by keeping critical entities in structured session state or re-injecting them into each specialist's context, not by hoping the raw transcript keeps everything forever. Bigger rate limits do not restore dropped turns. Safety filters are not inventing new policy numbers. Re-embedding the index does not put missing entities back into the conversation window the writing agent actually sees.
Full Explanation
Option A is correct. Fixed-size sliding windows discard oldest turns as conversations grow. When identity-bearing facts live only in those early natural-language turns, later agents lose entity continuity even if RAG still surfaces the right claim documents. The first remediation is architectural: store durable entities (policy number, claimant name, claim id) in multi-tier state and inject them into each agent prompt, or use summary-plus-entity memory that explicitly preserves identifiers. That targets the observed mismatch between successful retrieval and hallucinated identifiers in generation.
Option B is incorrect because throughput quotas affect how many tokens you may send per minute, not which historical turns the window retains. Raising rate limits does not prevent the window from dropping early content.
Option C is incorrect because content safety may block certain categories of harmful content; it does not systematically remove claim identifiers and then cause agents to invent replacement numbers. Disabling safety for internal agents is also a security regression unrelated to the root cause.
Option D is incorrect because embedding refresh can improve retrieval quality, but the stem states retrieval already finds the correct claim. The failure is in conversational context continuity for generation, not vector search precision.