Online deployments of agents generate massive execution traces, yet task‑specific verification and expert annotation are expensive to scale. We investigate how to distill these traces into reusable feedback without post‑hoc outcome labels, leveraging evidence of local progress, recovery actions, and unfinished requirements.
We introduce DENSE (Distilling Evidence from Nested Subtask Executions), which organizes such evidence into evidence‑grounded nested shortcut trees. The tree compresses redundant attempts, uses recovery evidence to reconcile conflicts across levels, summarizes completed branches and expands unresolved ones, linking existing progress to remaining obligations.
To evaluate feedback quality, we propose the REFIT protocol: under outcome blindness, the same initial trajectories are paired as sources, environments and model contexts are reset, and different models re‑attempt the same tasks, allowing a fair comparison of the derived feedback.
On the Terminal‑Bench 2.1 benchmark, DENSE achieves the highest strict pass rate among four recipient models, outperforming all non‑privileged feedback methods. Compared with the original executions, strict pass rate improves by 7.12‑15.64 percentage points, while observed recipient tokens in reruns drop by 19.0‑43.6%. Ablations with GPT‑5.5 confirm that nested subtask analysis, shortcut‑tree construction, and issue reconciliation are all essential.
These findings suggest that evidence‑grounded trajectory reuse can enable agents to self‑refine, reducing reliance on external supervision.
Review