Summary

SkillX is an automated framework for constructing reusable, plug-and-play skill knowledge bases for LLM agents from experience. It distills agent interactions into a structured three-level skill hierarchy (Planning, Functional, Atomic), enabling robust skill transfer and significantly improving task success and efficiency for agents, particularly enabling strong-to-weak transfer.

How It Works

SkillX employs a three-tier skill hierarchy: Planning Skills for high-level task organization, Functional Skills for reusable tool-based subroutines, and Atomic Skills for execution-oriented tool usage patterns. Its fully automated pipeline extracts, filters, and consolidates skills from agent trajectories. The system supports iterative refinement and exploratory expansion to proactively discover and improve skills, creating a transferable library that can be directly injected into different base agents without retraining. This structured abstraction offers greater reusability and generalization than raw trajectories or abstract reflections.

Quick Start & Requirements

The provided README text indicates "Installation" and "Quick Start" sections exist but does not detail specific commands, non-default prerequisites (e.g., Python versions, hardware), or estimated setup times. Links to official quick-start guides or documentation are not present in the provided text.

Highlighted Details

• Achieves ~10% absolute improvement for weaker base agents on multiple benchmarks.
• Consistently improves task success and execution efficiency on challenging benchmarks like AppWorld, BFCL-v3, and τ2-Bench.
• Demonstrates stronger transferability compared to trajectory-based, workflow-based, and memory-based baselines.
• Enables effective strong-to-weak transfer, where a skill library built by a stronger model can be used by weaker ones.

Maintenance & Community

The project acknowledges contributions from Ant Digital Technologies and Ant Group. It builds upon and adapts code from ReMe, AgentEvolver, A-MEM, AWM, and Expel. No specific community links (Discord, Slack) or roadmap details are provided in the text.

Licensing & Compatibility

The project is licensed under the MIT license, which is permissive and generally compatible with commercial use and closed-source linking.

Limitations & Caveats

The provided README focuses on the project's capabilities and benefits and does not explicitly mention any limitations, alpha status, known bugs, or unsupported platforms.