Skills vs Agents vs MCP

Agents are the workers. Skills are reusable task knowledge. MCP servers expose tools and data. AGENTS.md and CLAUDE.md give standing project instructions. Plugins, commands, hooks, and subagents package or enforce parts of that system.

These concepts are easy to blur because they all influence what an AI coding agent does. The useful way to separate them is by job: who reasons, what knowledge gets loaded, what external systems can be called, and what governance layer records or blocks actions.

In a governed AI engineering stack, the layers are complementary. A single agent task might read AGENTS.md, load a code-review skill, call an MCP server for pull request context, run a command, and hand the result to a security subagent for review.

The mistake is treating one layer as a replacement for every other layer. A skill does not make an agent safe. An MCP server does not teach the agent your review policy. A subagent does not remove the need for project context. Governance comes from designing the whole stack, not from naming one component.

The table below uses implementation-neutral language. Claude Code, Codex, OpenCode, OpenClaw, and other runtimes may support these concepts with different file paths, frontmatter fields, invocation models, or permission controls.

A prompt is the fastest way to steer a model in the current turn. It is also easy to lose, copy inconsistently, or forget to include when the next task starts. A skill is for the prompt that keeps coming back.

Use a skill when the workflow has a predictable shape: review a diff, write release notes, update documentation, verify a frontend change, or inspect a security-sensitive path. A skill can carry the instructions, examples, checklists, and scripts that make the workflow repeatable without pasting a long prompt every time.

For the category overview, start with what are agent skills.

Project instruction files describe the terrain: repository conventions, test commands, ownership boundaries, review process, style rules, and standing policies. They should be read before the agent does any work in that project.

Skills describe repeatable moves inside that terrain. A code-review skill might apply across many repositories. A repo's AGENTS.md explains the local rules that change how the skill should behave in this repository. The two are strongest together: the skill supplies the workflow, the project instructions supply the local constraints.

Skills tell the agent how to do something. MCP servers and tools let the agent do something outside the model: search a code index, fetch a Jira ticket, query a database, read docs, create a pull request, or call an internal API.

A skill can instruct the agent to use an MCP tool, but it should not be confused with that tool. The skill is the procedure. The MCP server is the external interface. The agent is the runtime that decides, under policy, whether and how to use both.

For MCP fundamentals, see what is Model Context Protocol.

These concepts often appear in the same product docs as skills, but they serve different jobs.

• Plugins package extensions so they can be distributed or enabled as a bundle.
• Commands give users an explicit trigger for a known workflow.
• Hooks run at event boundaries, which makes them useful for deterministic enforcement.
• Subagents delegate work to a specialized role or separate reasoning context.

A single plugin might include skills, commands, hooks, and MCP configuration. A command might invoke a skill. A hook might block a risky command before the skill can execute it. A subagent might use the same skill with a narrower role.

Pick the smallest layer that solves the problem. If a rule needs enforcement, a skill is too soft by itself. If a workflow needs reusable judgment, a shell script is too narrow by itself. If a task needs external data, prompt text alone will stale quickly.

Once these layers multiply, the real problem becomes visibility. Which instructions were active? Which skill loaded? Which MCP tool was called? Which subagent reviewed the diff? Which human approved the result?