HarnessBridge: Learnable Bidirectional Controller for LLM Agent Harness

HarnessBridge is introduced as a lightweight learnable bidirectional controller designed to enhance the agent-environment interface for long-horizon tasks in large language models (LLMs). It employs two bidirectional projections—observation projection for condensing raw trajectories into decision-relevant states, and action projection for transforming proposed actions into executable transitions. Trained via unified instruction tuning on a harness supervision dataset, HarnessBridge demonstrates competitive performance on Terminal-Bench 2.0 and SWE-bench Verified, achieving reduced token usage and trajectory length while generalizing effectively across model sizes, thereby offering a scalable solution for practitioners working with LLMs in complex environments.