Improving long‑horizon LLM agents can follow two avenues: evolve a harness around a frozen model, or train the model's weights. We let a self‑evolving harness first strengthen the system, then cross‑seed the evolved harness with the base model and with trained weights to see which gains the trained model retains and which still require runtime fixes.
The key is to read the agent's failure composition: labeling each failed trajectory by the first signal that fires separates failures into process failures (blocked calls, loops, exhausted step budgets) and content failures (a delivered plan that is poor). Harness evolution repairs process failures, and the behavior it instills can later be baked into the weights; content failures are what weight training targets.
On the DeepPlanning benchmark, a self‑evolving harness loop lifts Qwen3.5‑4B's held‑out score from 0.16 to 0.30 and Qwen3.5‑9B's from 0.32 to 0.44; for the 4B model, delivery rises from 55% to 90% while content failures remain for the weights to handle.
We train LoRA adapters on the evolved‑harness trajectories. Under the original harness they add +0.13 to held‑out tasks for both sizes; on 4B they stack with the harness to more than double the held‑out score, and on 9B the adapter alone matches the full evolution line, cutting content failures from a quarter of trajectories to one in twenty. A placebo adapter trained on answer‑shuffled trajectories falls below the base model.
The loop transfers to WebArena‑Lite (+0.09 on 117 unseen tasks), where the gain lives in what the model sees and adapters do not add further improvement.
The resulting diagnose‑then‑intervene rule is applied twice: read the failure composition to choose between harness and weights, then read what the accepted edits changed to decide which gains to train into the weights. Scores are four‑rollout means against fresh anchors, same‑night except where marked, across eight models from six families and two benchmarks.
Review: The paper presents a systematic way to combine runtime harness evolution with offline weight training by first diagnosing failure types. Experiments show that harness evolution quickly fixes process bottlenecks, while content‑level shortcomings still benefit from weight updates or LoRA adapters. This two‑stage diagnose‑intervene framework offers a practical roadmap for long‑term LLM optimization.