An Agent Skill is a folder containing a markdown file, usually named SKILL.md, that tells an agent how to do something. The agent reads the file's name and description up front, decides mid-task that the skill is relevant, and pulls the full body into context only then. Anthropic published the format as an open specification and shipped it in Claude Code and the Claude apps; the pattern spread to other harnesses quickly because it requires no new infrastructure. A skill is a file in a folder. That is the entire distribution mechanism, and it is why the ecosystem grew the way it did.
The usual mental model is that skills are how you give an agent knowledge it lacks, and the research does not support that reading. A study of 528 paired skill-and-task triples labelled 65.7% of skill mechanisms procedural anchoring, meaning they constrain how the agent acts, against 4.5% knowledge injection. Skills are closer to a house style guide than to a reference manual. They tell an agent to always run the linter before committing, to structure a report in a particular order, to prefer one library over another. That is real value, and it is not the same value as memory.
Progressive disclosure is the mechanism that makes a large library affordable. Metadata for every skill sits in context permanently, which is cheap; bodies load only when selected, which is expensive but rare. The design implies that library size is free, and at scale it is not. Selection is the bottleneck rather than authoring: as candidate pools grow from five skills to a hundred, measured execution-time precision on which skills actually get used falls from 29.6% to 3.3%. The problem is not how many skills you have, it is how confusable they are with each other.
Skills are also not memory, and conflating the two causes a specific failure. A skill file is static text an author wrote; memory is state a system maintains about a user and a world that keeps changing. Teams routinely route memory through skill files because both are text the agent reads, and it works for stable facts and fails for anything that has to be updated, contradicted, scoped to a person, or forgotten. Maximem Synap exists on the other side of that line, handling the state that a folder of markdown cannot.
The governance story is the least developed part of the ecosystem. A skill file carries the authority of a user instruction and passes through none of the review a user instruction would face: no signing, no provenance, no sandbox by default. A mining run across public repositories found 3.8 million SKILL.md files, and 50.5% of them are verbatim copies of another file, which is a supply chain with no supply chain controls attached to it.
Why It Matters
Skills are the first widely adopted way to change agent behaviour without touching the model or the harness, which makes them the natural first thing a team reaches for. Understanding what they are good at, and specifically that they anchor procedure rather than supply state, is what stops a team from building their memory layer out of markdown files and discovering the gap in production. The distinction also sets the boundary of the technology: skills shape the how, a memory layer maintains the what, and a serious agent needs both.
Example
A platform team writes a skill that tells their coding agent how to open a pull request in their house style: branch naming, a required test plan section, no direct commits to main. Every engineer's agent picks it up and the PRs become consistent overnight. The same team then tries to put customer context into a second skill file, one per account, and it degrades within a month. Accounts change owners, deals close, preferences reverse, and nobody remembers to edit the markdown. The first skill was procedure and it held. The second was state and it rotted.