Imagine your class is making a big art project together. One kid draws the shapes, one kid colors them in, and one kid cuts them out. That is a multi-agent workflow: a team of helpers, each doing one job. Now your teacher wants to add a new helper who writes the labels. She gives the new kid a name tag. The name tag says who they are and what they are good at. It also says which tools they can borrow from the supply box. That name tag is the agent profile. The teacher reads every name tag before she asks anyone for help. She looks for the right helper and hands them the task. Sometimes a helper also carries a special folder in their backpack. The folder holds step-by-step instructions for tricky jobs, like how to write in bubble letters. The helper pulls out the folder only when the task needs it. That folder is an agent skill. The new helper joins the group and starts working. Everyone still does their own job. The new helper does not take over. They only do the thing written on their name tag. But here is the danger. If the new helper writes a name tag that says I do everything, the teacher gets confused. She sends the wrong helper to do the wrong job. The label writer tries to cut paper. The cutter tries to write. The whole project slows down. A good agent profile is clear and specific so the right helper always gets the right task.
Both declare a specialist's identity, listed capabilities, and allowed tools before any work begins, so the orchestrator knows exactly who to call and for what.
Both are loaded on demand: the helper only pulls out the folder when the current task matches, just as Copilot only injects a SKILL.md when the prompt matches the skill's description.
Listing specific tools on the name tag limits what the helper can reach for, preventing a labeling specialist from accidentally grabbing scissors and overstepping their role.
The orchestrator reads each agent profile's description to route a task correctly, just as the teacher reads name tags to pick the right specialist rather than asking anyone at random.
Adding an agent to a workflow means writing a clear, specific name tag so the orchestrator always picks the right specialist for the right task.