A multi-agent claims workflow in Microsoft Foundry keeps a shared working context that accumulates every agent’s intermediate reasoning, tool payloads, and retrieved passages. After several long-running cases, later agents begin truncating mid-sentence because the context window is full, and critical claim identifiers from early turns are dropped first. Which approach best restores usable shared context without losing high-value entities?
Select an answer to reveal the explanation.
Short Explanation
The shared working memory is overflowing, so you need to compact it intelligently. Summarize low-value turns, keep structured entities like claim IDs, and preserve recent tool results before passing context to the next agent. Turning up temperature just makes the model guess, and that's dangerous in claims.
Full Explanation
The correct approach is to implement a context compaction pipeline that summarizes lower-priority turns while preserving structured entity fields and recent tool results before injecting context into the next agent. In multi-agent solutions, context accumulation across hops frequently exhausts model context windows. Deliberate compaction strategy—summarizing verbose intermediate reasoning while retaining structured entity fields (claim IDs, policy numbers), critical decisions, and recent tool results—keeps shared context both within budget and operationally complete before the next agent is invoked.
Increasing only the final agent’s temperature so it invents missing claim identifiers when context is truncated is incorrect because temperature only influences sampling randomness; it does not restore truncated identifiers and may increase hallucinated values when data is missing.
Disabling all tool calls so agents stop adding tool payloads and rely solely on the original user message is incorrect because tools often supply the authoritative evidence (policy lookups, claim status). Removing tools reduces noise but also removes grounding, and does not solve entity continuity across agents.
Restarting the conversation on every agent hop so each specialist begins with an empty context window is incorrect because it discards prior agent decisions and shared state, breaking multi-agent continuity by design.
Exam focus: diagnose context overflow as a management problem (compaction, entity preservation, selective retention), not a model-temperature or “start clean every time” problem.