Rethinking Agent Trajectory Compression: A Workflow-Aware Perspective
Summary
- Researchers proposed WAAC, a workflow-aware framework for agent trajectory compression designed to reduce context overhead in long-horizon agents.
- WAAC models successful historical trajectories to extract reusable compression experience and uses online backward dependency analysis to identify retention requirements.
- Experiments on PinchBench and Odysseys show WAAC achieves an overall score of 87.6 and a completion rate of 19.0% while reducing execution costs by 19.7% and 29.1%.
Why it matters
This research provides a more efficient approach to managing long-horizon agent trajectories, reducing context overhead and execution costs without discarding critical task information.