SkillWiki: A Living Knowledge Infrastructure for Agent Skills
SkillWiki is a novel infrastructure designed to address the lack of large-scale production and governance for agent skills by transforming heterogeneous knowledge into reusable, provenance-aware assets that enable the continuous co-evolution of knowledge, skills, and execution experience.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a brilliant, super-smart robot assistant. Right now, teaching this robot new tricks is a bit chaotic. You might give it a specific instruction for one task, then a different note for another, and over time, you end up with a messy pile of instructions that are hard to find, hard to fix, and impossible to update when things change.
The paper introduces SkillWiki, a solution to this mess. Think of SkillWiki not just as a storage box, but as a living, breathing library and factory combined for robot skills.
Here is how it works, using simple analogies:
1. The Problem: The "Messy Garage" vs. The "Organized Library"
Currently, if you want to add a new skill to an AI, it's like throwing a new tool into a crowded garage. You know it's in there somewhere, but you don't know exactly where, who made it, or if it's broken.
- Wikipedia is how we manage human knowledge: anyone can edit, but there's a history of changes, and everything is linked.
- GitHub is how we manage software code: it tracks every version, who changed what, and allows for testing before releasing.
- SkillWiki says: "AI skills need their own version of Wikipedia and GitHub." It treats skills not as static notes, but as living assets that need to be organized, tracked, and improved over time.
2. The Factory: Turning Raw Materials into Tools
SkillWiki acts like a factory that takes raw materials and turns them into polished tools.
- Raw Materials: The system accepts all kinds of messy inputs: chat logs, PDF manuals, code scripts, API lists, or even recordings of the robot failing at a task.
- The Process: Instead of just saving these files, SkillWiki's "workers" (automated agents) read them, extract the useful "how-to" parts, and turn them into a structured, reusable Skill.
- The Receipt: Crucially, every new skill comes with a "receipt" (provenance). You can always look back and see exactly which document or chat log the skill came from. If the original manual changes, you know which skills might need updating.
3. The Lifecycle: A Skill's Journey from Birth to Retirement
The paper describes a specific journey a skill takes, which is like a product going through a factory line:
- Raw Experience (S0): A raw log of something happening.
- Candidate (S1): The system guesses, "Hey, this looks like a useful skill."
- Draft (S2): A human or AI writes it up properly.
- Verified (S3): The skill is tested in a safe environment to make sure it actually works.
- Released (S4): It's officially available for the robot to use.
- Degraded (S5): If the skill starts failing (e.g., the website it uses changed), the system flags it as "sick."
- Deprecated (S6): It's marked as "outdated" so no one uses it anymore.
- Archived (S7): It's put in a museum for history, but kept safe in case we need to remember how it used to work.
4. The Self-Healing Mechanism
This is the most "sci-fi" part that the paper claims. SkillWiki doesn't just sit there; it watches the robot work.
- The Watchdog: If the robot tries to use a skill and fails (like trying to open a door that no longer exists), the system notices.
- The Doctor: It analyzes why it failed. Did the environment change? Is the code wrong?
- The Surgeon: It automatically proposes a fix or a new version of the skill. It runs a "Git-style" review (like a code update), checks the changes, and if they look good, it releases the new, fixed version while archiving the old, broken one.
5. The Demonstration
The authors built a working system to show this off. They showed that:
- They could take 125 different types of messy documents and turn 99 of them into working, verified skills.
- They could simulate a full lifecycle where a skill is born, used, breaks, gets fixed, gets a new version number, and eventually gets retired, all while keeping a perfect record of every single step.
Summary
In short, SkillWiki is an infrastructure that stops treating AI skills like static files in a folder. Instead, it treats them like living products in a managed ecosystem. It connects the "what" (the skill), the "where" (the source document), and the "how" (the execution history) so that AI agents can have a massive, organized, and self-correcting library of abilities that grows smarter and more reliable over time.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.