A multi-agent investigation workflow in Microsoft Agent Framework accumulates tool outputs and intermediate reasoning until agents hit context window limits and start dropping critical entity names. Which implementation strategy best manages context across agents?
Select an answer to reveal the explanation.
Short Explanation
You need deliberate context management: accumulate selectively, retrieve relevant spans, inject structured context into the next agent, and compact while preserving entities and decisions. That stops sliding-window amnesia and keeps case identifiers, parties, amounts, and open questions alive across long multi-agent runs.
Full Explanation
The correct strategy is selective accumulation, retrieval of relevant spans, structured injection into downstream agents, and compaction that preserves entities and decisions. Developers must manage context for single- and multi-agent systems through accumulation, retrieval, injection, and compaction. Structured handoffs preserve signal while controlling token growth. Pair compaction with evaluations that detect summary drift and entity loss before production. Make compaction loss-aware: preserve case identifiers, parties, amounts, and open questions, not generic short summaries. Disabling all tools avoids growth but destroys solution value. Unbounded concatenation of raw payloads and chain-of-thought causes overflow, noise, and diluted attention. Browser-only localStorage is not secure or reliable for multi-agent context.