SkillX: Automatically Constructing Skill Knowledge Bases for Agents
| Authors | Chenxi Wang et al. |
| Year | 2026 |
| HF Upvotes | 31 |
| arXiv | 2604.04804 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Learning from experience is critical for building capable large language model (LLM) agents, yet prevailing self-evolving paradigms remain inefficient: agents learn in isolation, repeatedly rediscover similar behaviors from limited experience, resulting in redundant exploration and poor generalization. To address this problem, we propose SkillX, a fully automated framework for constructing a plug-and-play skill knowledge base that can be reused across agents and environments. SkillX operates through a fully automated pipeline built on three synergistic innovations: (i) Multi-Level Skills Design, which distills raw trajectories into three-tiered hierarchy of strategic plans, functional skills, and atomic skills; (ii) Iterative Skills Refinement, which automatically revises skills based on execution feedback to continuously improve library quality; and (iii) Exploratory Skills Expansion, which proactively generates and validates novel skills to expand coverage beyond seed training data. Using a strong backbone agent (GLM-4.6), we automatically build a reusable skill library and evaluate its transferability on challenging long-horizon, user-interactive benchmarks, including AppWorld, BFCL-v3, and τ^2-Bench. Experiments show that SkillKB consistently improves task success and execution efficiency when plugged into weaker base agents, highlighting the importance of structured, hierarchical experience representations for generalizable agent learning. Our code will be publicly available soon at https://github.com/zjunlp/SkillX.
Engineering Breakdown
Plain English
SkillX is a framework that solves a critical problem in LLM agent development: agents currently learn in isolation and waste effort rediscovering the same skills repeatedly across different tasks and environments. The paper proposes a fully automated pipeline that constructs a reusable skill knowledge base organized in three hierarchical levels—strategic plans, functional skills, and atomic skills—that can be plugged into multiple agents. The key innovation is combining multi-level skill distillation with iterative refinement to enable agents to learn from each other's experiences and generalize across different tasks, dramatically reducing redundant exploration and improving sample efficiency.
Core Technical Contribution
The core technical novelty is the three-tiered skill hierarchy that automatically abstracts raw trajectories into different levels of granularity: atomic skills are primitive operations, functional skills are compositions of atomic skills that achieve specific goals, and strategic plans are high-level decision policies that orchestrate functional skills. This hierarchical abstraction allows skills learned in one environment or task to be transferable to others without manual engineering or domain expertise. The iterative refinement component automatically validates and improves skills based on real-world performance feedback, creating a self-improving system rather than a static skill library. This differs from prior approaches by (1) automating the entire skill extraction pipeline rather than requiring manual skill definition, (2) organizing skills hierarchically for composability and reuse, and (3) continuously refining skills rather than treating them as static once learned.
How It Works
SkillX operates through a pipeline with three main stages: First, trajectory collection aggregates experiences from multiple agents across different environments and tasks into a unified dataset. Second, multi-level skill distillation processes these trajectories bottom-up: atomic skills are extracted from primitive action sequences, functional skills are discovered by identifying recurring patterns and goal-oriented subsequences in trajectories, and strategic plans are learned by analyzing high-level decision sequences across trajectories. The skill hierarchies are represented as executable code or policies that new agents can directly invoke. Third, iterative refinement continuously improves the skill library by (a) testing skills on new tasks and environments, (b) identifying failures or inefficiencies, (c) refactoring or recomposing skills to fix issues, and (d) retraining or fine-tuning skill implementations based on performance metrics. The skill knowledge base becomes a centralized repository where any new agent can query and reuse skills rather than learning from scratch, enabling warm-start initialization and cross-task transfer.
Production Impact
In production, SkillX would dramatically reduce the cost and time required to deploy new agents across different tasks and environments. Instead of training each agent independently (which requires expensive exploration and potentially thousands of environment interactions), a new agent starts with pre-trained, validated skills from the shared knowledge base, reducing cold-start time from hours or days to minutes. Teams would operate a single skill management system rather than training isolated agents—this centralizes observability, enables shared debugging, and allows skills to improve continuously as more agents contribute trajectories. The main trade-off is upfront investment: you need infrastructure to aggregate trajectories, automated skill extraction pipelines, and continuous validation of skills before deploying them. Compute cost shifts from per-agent training to a centralized skill library maintenance cost, which is likely lower overall but requires careful monitoring of skill quality and versioning.
Limitations and When Not to Use This
The paper assumes trajectories contain sufficient signal to extract meaningful skills, but in sparse-reward environments or early-stage exploration, trajectory data may be too noisy or uninformative for reliable skill distillation. The hierarchical skill design requires defining what constitutes an atomic skill, functional skill, and strategic plan—this boundary may be domain-specific and isn't fully automated in the framework. The iterative refinement loop depends on having good validation metrics; if a skill appears successful in validation but fails in production on edge cases, the feedback loop may propagate the failure across all agents using that skill. The paper doesn't address skill forgetting or interference: as the skill library grows, managing versioning, deprecation, and ensuring new skills don't break existing agent behaviors becomes operationally complex. Generalization across very different domains (e.g., from text-based agents to robotic agents) is not thoroughly validated, so the reusability claim may only hold within narrow task families.
Research Context
SkillX builds on the broader trend of moving LLM agents from single-task learning to multi-task and continual learning paradigms, similar to work in curriculum learning and hierarchical reinforcement learning but applied to the LLM agent setting. It addresses a pain point identified in recent work on self-improving agents and in-context learning, where agents waste computation rediscovering solutions. The framework is positioned as a middle ground between fully manual skill libraries (which don't scale) and pure end-to-end learning (which is inefficient). This opens research directions in skill compositionality, semantic skill discovery, and cross-domain skill transfer—questions like how to automatically discover meaningful skill abstractions, how to detect when skills from different domains are semantically equivalent, and how to merge or adapt skills learned in one environment for use in another.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::