HarnessBridge replaces hand-engineered agent harnesses with a learnable controller
HarnessBridge is a lightweight, trainable harness controller that parameterizes the agent–environment interface as a bidirectional projection, matching or surpassing specialized hand-crafted harnesses on Terminal-Bench 2.0 and SWE-bench Verified while reducing token usage and trajectory length.
Score breakdown
HarnessBridge replaces the manual engineering bottleneck in LLM agent harness design with an end-to-end trainable module, reducing token usage and trajectory length while maintaining competitive benchmark performance.
- 01HarnessBridge is a lightweight, learnable harness controller that parameterizes the agent–environment interface as a bidirectional projection.
- 02It learns two projections: an observation projection (distilling raw trajectories into compact states) and an action projection (converting actions into executable transitions or rejections).
- 03The model is trained on a harness supervision dataset via unified instruction tuning.
Xiaoxuan Wang, Haixin Wang, and Alexander Taylor present HarnessBridge, a lightweight learnable harness controller designed to address a key bottleneck in LLM agent deployments: the harness that mediates agent–environment interaction. Existing harnesses are largely hand-engineered, which makes them difficult to scale as task trajectories grow longer and interactions become more complex. HarnessBridge proposes to generate this harness component via a trainable plug-in module that can be optimized end-to-end.
The model is trained on a harness supervision dataset using unified instruction tuning.
The core of HarnessBridge is a bidirectional projection framework. The observation projection distills raw trajectories into compact, decision-relevant states, while the action projection converts proposed actions into either executable transitions or trajectory-grounded rejections. The model is trained on a harness supervision dataset using unified instruction tuning.
On the Terminal-Bench 2.0 and SWE-bench Verified benchmarks, HarnessBridge matches or surpasses strong specialized harnesses while substantially reducing both token usage and trajectory length. Notably, the system generalizes from smaller generator models to larger commercial models, suggesting the learned harness controller transfers across model scales.
Key facts
- 01HarnessBridge is a lightweight, learnable harness controller that parameterizes the agent–environment interface as a bidirectional projection.
- 02It learns two projections: an observation projection (distilling raw trajectories into compact states) and an action projection (converting actions into executable transitions or rejections).
- 03The model is trained on a harness supervision dataset via unified instruction tuning.
- 04Evaluated on Terminal-Bench 2.0 and SWE-bench Verified benchmarks.
- 05HarnessBridge matches or surpasses strong specialized harnesses on those benchmarks.
- 06It substantially reduces token usage and trajectory length compared to existing harnesses.
- 07The controller generalizes from smaller generator models to larger commercial models.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 12, 2026 · 10:05 UTC. How this works →