Stanford paper: harness edits fix agent process failures, weight training fixes poor plans

A Stanford-affiliated paper says agent failures split into process failures, such as loops and exhausted step budgets, and content failures, where the agent delivers a poor plan.

On a travel-planning benchmark, harness evolution raised Qwen3.5-4B's held-out score from 0.16 to 0.30 and plan delivery from 55% to 90%, but did not reduce poor plans. A LoRA adapter cut poor plans for Qwen3.5-9B from 28% to 5% of held-out runs, according to a post by Rohan Paul on X summarizing the paper.