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?

AI-500 development objectives cover context management for single and multi-agent systems: accumulation, retrieval, injection, and compaction. Production orchestrators should maintain a working set—recent turns, pinned facts, retrieved memories—not the entire transcript. Compaction must avoid summary drift and entity loss (ticket IDs, account numbers, prior approvals). Option B fails at scale. Option C induces sliding-window amnesia. Option D is not an engineering fix. Combine structured state (case object) with summarized narrative and re-retrieve details from memory stores when needed.