Tencent ContextPilot Trains Agents to Manage Context

The important idea is treating context management as a learned control problem, not a fixed summarization layer. It could make agents more efficient, though branching rollouts add training complexity and the evidence remains limited to research benchmarks.

• –Adds planning, structured long-term memory, compression, summarization, and searchable history folding.
• –Uses context and entropy changes to focus exploration on high-impact editing decisions.
• –Averages rewards across downstream branches instead of assigning one final score to every edit.
• –Reports BrowseComp+ accuracy rising from 63.49% to 80.96% with the full toolset.
• –Releases code and model checkpoints, but leaves coding and GUI-agent evaluation for future work.