Agents

Every agent is a folder containing an agent.md. The folder’s position in the tree defines its place in the org chart — its subfolders (each with their own agent.md) are its direct reports.

Frontmatter

---
name: Acme CEO
role: ceo
model: opus
autonomy: orchestrated
budget:
  usd: 20
  turns: 40
---
You are the CEO of Acme. You set direction for the growth and research
teams and make the final call on anything they escalate to you.

Field	Type	Notes
`name`	string	Display name
`role`	string	A short label — used in logs and the console
`model`	`opus` \| `sonnet` \| `haiku`	Which model this agent runs on
`autonomy`	`orchestrated` \| `bounded`	See below
`budget`	`{ usd?, turns? }`	A ceiling on spend and/or turns for this agent

`autonomy`

orchestrated — this agent plans and dispatches tasks to its subordinates as part of a process, rather than doing the work itself. Typically used for team leads and managers.
bounded — this agent does the work directly, within its budget and tool grants, without further delegation.

The system prompt

Everything in the markdown body below the frontmatter is the agent’s system prompt. There’s no required structure — write it the way you’d brief a person taking on the role.

Tool grants

An agent’s tool access isn’t declared in agent.md — it’s declared separately in the agent’s own tools.json. See Tools & policies.

Next