During a long multi-agent troubleshooting session, the orchestrator’s context grows with tool traces and subagent reports until calls fail with context-length errors. You need continuity of key entities without sending the entire history every turn. What should you implement?
Select an answer to reveal the explanation.
Short Explanation
Implement context management with selective accumulation, retrieval of relevant prior turns, injection of only needed snippets, and compaction/summarization of older trace content while preserving critical entities. You keep continuity without blowing context windows. Full raw history, wiping everything, or disabling tools just breaks the session.
Full Explanation
Context management with selective accumulation, retrieval of relevant prior turns, injection of only needed snippets, and compaction/summarization of older trace content while preserving critical entities is correct. AI-500 covers accumulation, retrieval, injection, and compaction for single and multi-agent systems. Orchestrators should maintain a working set—recent turns, pinned facts, retrieved memories—not the entire transcript. Compaction must avoid summary drift and entity loss, such as ticket IDs, account numbers, and prior approvals. Always concatenating the full raw history of every agent and tool fails at scale. Clearing all context after every tool call, including the user's original goal, induces sliding-window amnesia. Disabling tools so context cannot grow is not an engineering fix. Combine structured state, such as a case object, with summarized narrative and re-retrieve details from memory stores when needed.