arXivPaperNeeds Review

Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

Reusable agent skills transferred better when induced from subtasks and stored as text. Whole-task skills often hurt versus no memory, so retrieval alone is not evidence that a memory helps.

arXiv · Aug 20, 2026

Open Source Open Markdown Open JSON

Source Summary

The study compares task-level against subtask-level skill induction and text against code formats. **Task-level skills often fell below the no-memory baseline**, while **subtask-level skills improved performance on average**; text transferred better than code.

Practical Implication

Build agent memory from small, reusable procedures and prefer textual guidance unless execution requires code. Before retrieval, score candidates for both task specificity and breadth of relevance rather than assuming any prior solution will help.

Agent-Ready Context

The study compares task-level against subtask-level skill induction and text against code formats. **Task-level skills often fell below the no-memory baseline**, while **subtask-level skills improved performance on average**; text transferred better than code.

Build agent memory from small, reusable procedures and prefer textual guidance unless execution requires code. Before retrieval, score candidates for both task specificity and breadth of relevance rather than assuming any prior solution will help.

Neither specificity nor abstractness predicted outcomes alone. Their combined **skill utility score** correlated with transfer performance and requires only skills plus task descriptions, but the supplied material gives no tasks, models, or effect sizes for judging generality.

Connected Context · Feed7 Judgment

This makes accumulated agent learning more selective than the prior “capture every lesson” pattern: reusable memory should be decomposed into small textual procedures and retrieved by joint specificity and breadth. It also gives a structural explanation for skill regressions—whole-task solutions can transfer worse than no memory—while leaving the proposed utility score’s generality unresolved without task and effect-size details.

The Regression Tax: Decomposing Why Skills Help and Hurt LLM AgentsIt reinforces the documented regression tax and narrows a likely cause: task-level skills often hurt, whereas smaller subtask procedures transfer better on average.SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient DescentSkillProx supplies the maintenance consequence of this result: candidate procedures should be tested for transfer, rolled back on regression, and pruned when utility is low.The Era of Compound Engineering — Kieran Klaassen, Every/CoraIt qualifies compound engineering’s lesson-capture strategy by showing that persistence alone is insufficient; lessons need subtask-level decomposition and relevance-aware retrieval.EveryInc/compound-engineering-pluginThe plugin is a concrete repository workflow for durable skills, while this study suggests its captured plans and solved-problem notes should be distilled into small textual procedures rather than reused wholesale.

Context Map

agent#skills#agent-memory#agent-reliability

Uncertainty

Neither specificity nor abstractness predicted outcomes alone. Their combined **skill utility score** correlated with transfer performance and requires only skills plus task descriptions, but the supplied material gives no tasks, models, or effect sizes for judging generality.