Agents
Building agents that patch other agents. This is an interesting approach for self-improving agents that leverages agent outputs. If you run …
Building agents that patch other agents. This is an interesting approach for self-improving agents that leverages agent outputs. If you run agents in production, you already have the training data for
Building agents that patch other agents. This is an interesting approach for self-improving agents that leverages agent outputs. If you run agents in production, you already have the training data for this. Every deployed agent accumulates failure trajectories and then does nothing with them. Harness-R1 turns those failures into runtime patches by post-training a dedicated 9B harness engineer with online RL. The engineer converts batches of target-agent failures into validated executable patches. Fresh reruns of the frozen target on the same batch supply outcome rewards. Only the engineer updates, so the target never drifts under the reward signal. Across WebShop, ALFWorld, and DBBench, vanilla Qwen3.5-9B goes from 44.3% to 53.6%. After the target itself is fine-tuned, a target-specific engineer lifts the average further from 59.2% to 64.2%. Paper: https://arxiv.org/abs/2608.02276 Track more trending AI papers in our academy: https://academy.dair.ai/
Source: DAIR.AI (X) | 2026-08-04