A research orchestrator and three specialist agents collaborate on long investigations. After about twenty turns, specialists begin contradicting earlier findings because the shared working context is truncated mid-entity. You need a context-management approach that preserves entity continuity across agents without exceeding model context limits. What should you implement?

The correct approach is context accumulation with periodic compaction into structured summaries plus selective retrieval of entity-critical facts before specialist handoffs. The exam expects you to design context management for single agents and between agents: accumulation, retrieval, injection, and compaction. For long multi-agent investigations, raw history grows faster than any model window. Compaction produces structured digests; selective retrieval rehydrates entities, constraints, and prior decisions at specialist handoffs so continuity survives truncation.

The approach that disables all memory and restarts a new chat for every specialist turn is wrong because wiping memory each turn maximizes isolation but guarantees specialists cannot share durable investigation state.

The approach that always passes the full raw transcript to every specialist regardless of length is wrong because unbounded full transcripts cause hard truncations, cost spikes, and attention dilution—exactly the failure mode described.

The approach that increases max_tokens on completions only and leaves context windows unchanged is wrong because completion max_tokens controls output size, not the input context budget that holds prior turns and shared state.

Practical pattern: compact → store structured team state → retrieve entity packs → inject into the next agent’s working context.