ContextPilot: Proactive Context Management via Fine-grained RL

Tencent Youtu Lab with Tsinghua and Shanghai AI Lab (co-first authors Zhuoshi Pan and Qizhi Pei) trains agents to manage their own working context on long-horizon tasks instead of letting interaction history grow unboundedly. Prior proactive context-management agents were limited to search/delete/summarize tools, explored context actions uniformly, and assigned one trajectory-level reward to every intermediate edit. ContextPilot widens the toolset (global planning, long-term memory, adaptive compression) and trains it with fine-grained RL that credits individual context-editing actions by their heterogeneous impact on outcomes.

Ships as post-trained Qwen3 checkpoints — ContextPilot-14B, -8B, and -E4B — released on HuggingFace on August 31, 2026 alongside code and a live demo. Squarely in the agent-harness/context-engineering line the index tracks (cf. active inference as context acquisition and WikiSkill's persistent-knowledge layer).