AI agent tests suggest harness edits can address stalls while fine-tuning reduces poor plans

A user says a Stanford paper sorted failed travel-planning runs into process failures, such as loops and exhausted step budgets, and content failures that produced poor plans. They report that harness edits raised Qwen3.5-4B from 0.16 to 0.30 on held-out tasks and increased plan delivery from 55% to 90%, but did not reduce the share of poor plans. A LoRA fine-tuning adapter cut poor plans from 28% to 5% of Qwen3.5-9B’s held-out runs.