Skip to main content

Agents

The second layer. An agent is a role — a name, a model, and a default set of skills — that a team's steps reuse by name. Agents live as native .claude/agents/<name>.md files (local to a project or global under ~/.claude/agents/), so they're the same primitive Claude Code itself uses.

Create an agent​

By an agent (Claude-drafted) — describe the role, Claude drafts it and auto-selects the skills that fit:

mmt new agent coder "implements a planned change cleanly and opens a PR"

Manually (author it yourself) — --ui opens the interactive Ink editor: description, model, and default-skills checkboxes:

mmt new agent coder --ui

Edit an agent later​

mmt edit agent coder "also run the test suite before opening the PR"   # bare — describe the change, Claude rewrites it
mmt edit agent coder --ui                                              # --ui — edit description / model / skills yourself

Worked example: coder​

The coder agent, at .claude/agents/coder.md:

---
name: coder
description: Implements a planned change cleanly, follows repo conventions, and opens/updates a PR.
model: sonnet
skills: [github-pr]
---

# coder

Implement the plan for the current step, following the repo's conventions and CLAUDE.md. Keep the change minimal and focused. When opening or updating a pull request, apply the github-pr skill.

Frontmatter carries name, description, an optional model, and the default skills: the agent plugs in; the body is the role's instructions. A team step can add more skills on top of these defaults (see Teams).

Where agents live​

Same two scopes as teams and skills:

  • local — ./.claude/agents/<name>.md, committed with the project.
  • global — ~/.claude/agents/<name>.md, available from any project.

List what you have:

mmt list agents          # alias: mmt agents