PASE Framework: Reconstructing Agent Skill Evolution with Executable Programs

PASE splits Agent skills into editable programs and natural language descriptions, enabling patch-based targeted evolution and outperforming baselines by over 2.8%.
Existing LLM Agent skills are mostly represented as natural language documents, mixing procedural logic with descriptive knowledge, which makes maintenance difficult and targeted updates infeasible. PASE proposes a program-aided patchable skill evolution framework that separates reusable procedural logic into editable programs, leaving natural language for descriptive knowledge only, enabling targeted modification of executable components similar to "patching." Experiments across multiple models and benchmarks show PASE consistently outperforms all baselines by over 2.8% and performs better in evolution stability and generalization. The article argues this marks Agent skill evolution moving from "writing documents" toward an engineering route of "writing programs + patching," with maintainability as important as performance, though programmatic representation introduces new challenges in runtime environment and cross-model compatibility.
When Agent Skills Learn to "Patch"
For LLM Agents to solve complex interactive tasks, skills are key—they encode reusable knowledge without modifying model parameters. Recently, many studies have attempted to distill experience from Agent execution trajectories and then iteratively refine skills based on downstream task feedback. However, existing methods mainly represent skills as natural language documents, where procedural logic and descriptive knowledge are intertwined, making maintenance and evolution difficult.
PASE (Program-Aided Patchable Skill Evolution) proposes a new path: associating skills with executable programs and iteratively evolving skills through program-aided means. The paper claims consistent improvements over all baselines across multiple models and benchmarks, achieving gains exceeding 2.8%.

Pain Points of Natural Language Skills
Writing skills as natural language documents is intuitive but structurally flawed. Procedural logic (what to do first, what next, how to fall back on errors) and descriptive knowledge (concept explanations, constraints) are mixed in the same text. Once a step needs correction, the entire description often has to be rewritten, potentially introducing unrelated changes.
The bigger problem is the difficulty of targeted updates. Experience distilled from Agent execution trajectories often consists of reusable operation sequences. Expressing them in natural language loses precision and makes them prone to different interpretations by models. The more complex the skill, the higher the maintenance cost of natural language documents, and the more unstable the evolution process.
LLM Agent skill distillation typically refers to extracting reusable experience from an Agent's historical task execution trajectories and solidifying it into a skill representation for direct use in subsequent tasks. This process requires no model parameter updates, making it low-cost and continuously accumulative. However, distilled experience often contains many specific operation sequences and contextual constraints. Recording it only in natural language makes it hard to precisely reproduce execution logic and inconvenient to modify as needed later. PASE targets exactly this representation bottleneck—separating reusable procedural knowledge from natural language documents and entrusting it to executable programs.
PASE's Core Idea: Separating Programs from Descriptions
PASE's approach splits a skill into two parts: editable programs carry reusable procedural logic, while natural language retains descriptive knowledge. This enables targeted updates to executable components, eliminating the need to repeatedly revise natural language skill documents.
This is equivalent to introducing a "patch" mechanism for skills: when downstream task feedback indicates that a certain step needs adjustment, you simply modify the corresponding program module while keeping the rest unchanged. The paper describes this as a program-aided framework, where skill evolution iterates with programs as the anchor.
Program-Aided here means that the core procedural logic of a skill is represented by executable program modules rather than plain text descriptions. Natural language still handles explaining concepts, constraints, and applicability conditions, but operation sequences such as "which step to execute first, how conditional branches proceed, how exceptions fall back" are encoded as programs. Thus, when downstream feedback requires adjusting a certain step, you only need to locate and modify the corresponding program module, forming a targeted update akin to a software patch without rewriting the entire skill document. Programmatic representation also makes skill behavior more deterministic and easier for different models to execute consistently, but it also requires skills to remain compatible with the runtime environment and interface conventions.
Experimental Performance and Stability
The paper conducted experiments across multiple models and benchmarks. PASE consistently outperforms all baselines, with gains exceeding 2.8%. Beyond absolute performance, further analysis shows that PASE performs better in evolution stability and generalization ability—which precisely correspond to the structural weaknesses of natural language approaches: drift caused by repeated revisions and overfitting of descriptive knowledge to specific tasks.
The authors emphasize that programmatic knowledge plays a key role in skill evolution. Data and code will be released on GitHub.
Evolution stability measures whether skill performance continues to improve rather than fluctuating violently or degrading across multiple rounds of iterative updates. Natural language skills are prone to introducing unrelated changes during repeated revisions, leading to skill drift; programmatic representation, through modular patching, can confine each modification to a local scope and reduce unintended side effects. Generalization ability refers to how skills perform when transferred to unseen tasks. If descriptive knowledge is overly tailored to specific tasks, it easily overfits; procedural logic carried by programs is more abstract and reusable, thus often performing more stably on new tasks. PASE's advantages in these two analyses are exactly the structural benefits brought by its representation approach.
Why It's Worth Attention
Agent skill evolution is moving from "writing documents" toward an engineering路线 of "writing programs + patching." PASE's value lies not only in that 2.8% improvement but more in placing skill maintainability on par with performance. For Agent systems that require long-term iteration and multi-task reuse, a skill representation that supports targeted patching may be more practical than one-off distilled natural language experience.
Of course, programmatic representation also brings new problems: program executability depends on the runtime environment and interface conventions, and cross-model compatibility requires additional design. The paper's experiments cover multiple models, but specific mechanism details remain to be verified after the code is open-sourced.
One-line Summary
PASE splits Agent skills into editable programs and natural language descriptions, enabling targeted skill evolution via patching, and consistently outperforms baselines by over 2.8%.
Paragraph Summary
Existing LLM Agent skills are mostly represented as natural language documents, where procedural logic and descriptive knowledge are mixed, leading to maintenance difficulties and the inability to perform targeted updates. PASE proposes a program-aided patchable skill evolution framework that separates reusable procedural logic into editable programs while natural language retains only descriptive knowledge. This supports targeted modification of executable components, forming an evolution mechanism akin to "patching." Experiments across multiple models and benchmarks show that PASE consistently outperforms all baselines with gains exceeding 2.8%, and performs better in evolution stability and generalization ability. The article argues that this marks Agent skill evolution moving from "writing documents" toward an engineering route of "writing programs + patching," where maintainability is as important as performance. However, programmatic representation also introduces new challenges regarding runtime environments and cross-model compatibility.
Related articles

Gemini 4 Neon Leak: A Potential Strong Contender Against Opus 5.5
A Hacker News thread suggests Gemini 4 Neon looks like an Opus 5.5 competitor. This article unpacks the product positioning behind this community signal, shifts in the top-model competitive landscape, and what it means for developers' model selection strategy.

Claude Fable 5.2's Self-Authored Short Film 'Fable' Sparks Discussion
A short film 'Fable,' credited to Claude Fable 5.2 for script, visuals, and directing, appears on Hacker News. This article analyzes the significance of full-pipeline AI creation, community reactions, and what it means for content creators.

Anthropic IPO Pricing: A Signal of AI Giants Going Public
Anthropic's IPO pricing draws attention. This article analyzes its potential impact on AI industry valuation, commercialization paths, and the competitive landscape, and highlights public-market expectations and challenges for AI companies.