Unified experience representation for agent skills

Determine which unified representation of agent experience achieves both plug-and-play usability and consistent effectiveness across diverse and complex tool-use scenarios for large language model agents.

Background

The paper surveys three major paradigms for encoding agent experience—case-based trajectories, strategy-based insights/workflows, and skill-based abstractions—and notes that different representations trade off transferability, retrieval efficiency, and executability.

Despite proposing a hybrid approach (high-level planning plus textual skills), the authors explicitly state that it remains unclear which unified experience representation would be broadly pluggable and consistently effective across heterogeneous, tool-intensive environments.

References

However, it remains unclear which unified experience representation is both easily pluggable and consistently effective, especially in diverse and complex agentic tool-use scenarios.

— SkillX: Automatically Constructing Skill Knowledge Bases for Agents  (2604.04804 - Wang et al., 6 Apr 2026) in Related Work, Encoding For Agent Experience

Distillation. The resulting trajectory runs to hundreds of turns interleaving correct and mistaken steps; deciding which to keep and how to compress a long, noisy demonstration into a clean, reusable skill is itself unsolved.

— Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design  (2609.22086 - Du et al., 18 Sep 2026) in Appendix, Section “Personalize: Where Retrieval-and-Injection Is Not Enough,” paragraph “Beyond the base tile: limits of the skill-based route”

If independently built agents are to be identified and composed at different stages of a trajectory, rather than relying on infeasible exhaustive profiling of every possible combination of agents, a compositional specification is needed, which provides a protocol for individual agents to expose a machine-processable interface not only of what they can do, but also of the tools, messages, and interactions through which their capabilities can be composed. The learnware principle identifies that interface, though a complete specification and reuse theory for general agents remains open.

— Learnware and AI Model Management System  (2609.11656 - Zhou, 10 Sep 2026) in Section 4.2, subsection “From identification to orchestration”

A simple execution log may prevent exact repetition, but transferring lessons across related workflows, models, and harnesses remains an open problem.

— Towards Efficient HPC Systems for Agents: Challenges and Opportunities  (2609.38723 - Zheng et al., 30 Sep 2026) in Section Agent Memory, subsection “Opportunities: A Hierarchical, Validated Memory Store”

How can transferred capabilities remain reliable and executable across heterogeneous agent environments?

— Federated Agent Optimization  (2610.01195 - Yang et al., 1 Oct 2026) in Section 7, Future Works, item 182

Whether our findings extend to such abstraction-based methods remains an open question.

— Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context Reinforcement Learning  (2610.11152 - Kwon et al., 8 Oct 2026) in Appendix, Section Extended Discussion of Limitations, paragraph “Scope: direct ICRL only”