ContextPilot trains agents to edit their own working context with per-action RL credit instead of one trajectory-level reward

arXiv 2608.28476·medium signalXBlueskyLinkedInCopy link

arXiv 2608.28476 (submitted 2026-08-28, from Tencent) targets three gaps in proactive context management: toolsets limited to search, delete and summarize with no global planning or adaptive compression; exploration that treats every context-editing action as equally consequential; and coarse credit assignment that smears the final trajectory reward across all intermediate edits. The framework expands the context-editing toolset and applies fine-grained RL credit to those edits specifically. It is the direct alternative to the summarize-and-truncate compaction most agent harnesses ship today.