Agent skills and plugins: How coding agents learn a new API without a human

An agent skill is a folder of instructions that a coding agent reads when a task calls for it. A plugin is a package that can bundle skills and other add-ons so they install in one step. Together they let an AI agent pick up a new service, such as a video API, without a person walking it through the documentation.

Picture hiring a contractor and handing them a binder. The cover lists what each section is for. They open a chapter only when the job needs it. A skill works the same way for an agent.

A folder with a cover sheet, opened only when needed

According to Anthropic’s engineering team, a skill is an “organized folder of instructions, scripts, and resources” that an agent can discover and load dynamically. The core file is SKILL.md. It starts with a short header holding two required fields, name and description, followed by the written instructions.

At startup, the agent loads only the name and description of each installed skill. If a task matches a description, the agent reads the full SKILL.md. If the instructions point to other files, such as a form-filling guide, the agent opens those only when needed.

Why loading in stages keeps the agent focused

An agent can hold only so much text in its working memory, called the context window. Pasting every manual into it would crowd out the actual task. Anthropic describes the staged approach as progressive disclosure: metadata first, the body second, linked files last.

The result is that an agent can know about many skills while paying for only the few it uses. It also makes the agent easier to steer, because each skill states plainly when it applies.

Skills can also carry code. Anthropic’s example is a PDF skill with a script that extracts form fields. The agent runs the script without reading the script or the PDF into context. Because code is deterministic, the output is repeatable, which is harder to promise when a model improvises the same steps in prose.

From a Claude feature to a shared format

Anthropic described Agent Skills in an engineering post on October 16, 2025. According to an update on that page, the company published Agent Skills as an open standard on December 18, 2025, so the same folder can work across different products.

Skills are related to, but different from, the Model Context Protocol (MCP). MCP connects an agent to live tools and data, such as an API endpoint. A skill tells the agent how to do a kind of work. In short, MCP supplies the reach and a skill supplies the know-how, and a vendor can offer both.

Plugins: The delivery truck for skills

A skill is just files, so something has to get it onto a developer’s machine. That is the job of a plugin. Cursor’s documentation describes a plugin as a distributable bundle that can package rules, skills, agents, commands, MCP servers and hooks. Hooks are scripts that run when a given event happens.

Cursor supports two formats. Agent Plugins follow an open standard, use a plugin.json manifest at the root, and package skills and MCP servers. Cursor Plugins use a .cursor-plugin/plugin.json manifest and can add rules, agents, commands and hooks.

Installing takes a few clicks: open Customize, pick the plugin, choose project or user scope. Cursor says each plugin in its marketplace is manually reviewed, must be open source, and that every update is reviewed as well. Teams and Enterprise plans can also run private team marketplaces.

Where it shows up today: Runway’s push into coding agents

Runway, the AI video company, has released a plugin for Cursor. In its announcement on X, the company said one install gives Cursor everything it needs to build Runway’s API into an app, as we reported in Cursor gets a Runway plugin, days after Codex got the same one.

That post does not say what the plugin contains, so it is unknown whether it ships skills, an MCP connection, or both. The pattern is clear, though. Instead of a developer reading API documentation, the agent receives packaged know-how and writes the integration itself. Runway’s claim that this works is the company’s own and has not been independently verified.

Limits and risks to weigh

Skills are instructions, and instructions can be wrong or hostile. Anthropic warns that a malicious skill could introduce vulnerabilities or direct an agent to exfiltrate data. Its advice is to install skills from trusted sources and, for less trusted ones, to audit bundled files, code dependencies and any instructions to connect to external networks.

A skill also does not guarantee good output. The agent still decides how to apply the guidance, so generated code needs the same review as any other code. Developers should check what an integration does with API keys and how much it can spend.

Some features depend on the product. Cursor’s documentation says rules, agents, commands and hooks work only in its own plugin format, and that running cloud agents keep the plugin versions they started with until a new agent is started. Plugin manifests also differ between tools, so a vendor may need to maintain several packages for one service.

What to watch next

Three things are worth following. First, whether the open standards for skills and plugins keep converging, so one package serves every agent. Second, how marketplaces vet submissions as the number of plugins grows. Third, whether more API providers, following Runway, treat an official plugin as a standard part of a developer launch.