Learning Agent-Compatible Context Management for Long-Horizon Tasks

In three linesAdaCoM, an external LLM system, manages context for frozen LLM agents via reinforcement learning on long-horizon tasks (web search, deep research). Learned strategies reveal a Fidelity-Reliability Trade-off: high-performing agents benefit from higher-fidelity context preservation, while lower-performing agents require aggressive compression.

Summary generated by Claude — human-verified