Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - Harness-R1 post-trains a dedicated 9B "harness engineer" model with online RL to edit an agent's executable runtime harness (context construction, tool mediation, action validation, execution recovery) based on failure trajectories, rather than updating the agent's own weights. It matters because it treats the scaffolding around an LLM agent as a learnable, self-improving component.
- The engineer converts batches of target-agent failures into validated executable patches; fresh same-batch reruns of the frozen target supply outcome rewards, so only the engineer is updated — edits are optimized for realized task success instead of being proposed by a fixed editor.
- Training pipeline: cold-start supervised fine-tuning initializes the editing policy, followed by online group-relative policy optimization (GRPO).
- On WebShop, ALFWorld, and DBBench, vanilla Qwen3.5-9B success rises from 44.3% to 53.6% (+9.3 pts).
- Gains persist after the target agent itself is fine-tuned: a target-specific engineer lifts the average from 59.2% to 64.2% (+5.0 pts), which the authors argue points toward co-evolving harness engineer and target agent.
Sources (1)
Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories
TL;DR - Harness-R1 post-trains a dedicated 9B "harness engineer" model with online RL to edit an agent's executable runtime harness (context construction, tool mediation, action validation, execution recovery) based on failure trajectories, rather than updating the agent's own weights. It matters because it treats the scaffolding around an LLM agent as a learnable, self-improving component.
- The engineer converts batches of target-agent failures into validated executable patches; fresh same-batch reruns of the frozen target supply outcome rewards, so only the engineer is updated — edits are optimized for realized task success instead of being proposed by a fixed editor.
- Training pipeline: cold-start supervised fine-tuning initializes the editing policy, followed by online group-relative policy optimization (GRPO).
- On WebShop, ALFWorld, and DBBench, vanilla Qwen3.5-9B success rises from 44.3% to 53.6% (+9.3 pts).
- Gains persist after the target agent itself is fine-tuned: a target-specific engineer lifts the average from 59.2% to 64.2% (+5.0 pts), which the authors argue points toward co-evolving harness engineer and target agent.