Agents & skills

An agent is a system that plans and acts with some autonomy; a skill is a reusable, versionable unit of instructions and workflow. Below is a shared vocabulary for when to lean on agents vs when to codify skills.

What is an agent

A goal-driven program that understands context, decomposes work, and calls tools (editor, shell, browser, MCP, etc.), adapting across steps. Best for open-ended exploration, triage, and integration across sources.

What is a skill

Often a SKILL.md (or similar) describing intent, I/O, guardrails, and checklists so different agents behave consistently. Best for repeatable process, reviewability, compliance trails, and team alignment.

Favor the agent when

Goals shift often; you need judgment (incidents, research, conversational clarification); you rely on live tool feedback; or the task is one-off exploration—run it first, then decide whether to extract a skill.

Favor a skill when

The same class of task repeats (review checklists, ship gates, license scans); deliverables have a stable shape (YAML, tables, report sections); you need reuse and versioning; or the workflow is tied to CI and repo conventions.

In practice they combine: agents orchestrate execution; skills supply pluggable step templates. Browse the full skill index; for business-style “when to trigger, when done” narratives see more cases.