Module 6 Agent Skill

Module 6 — Deploying with the Deepagents CLI

Lesson Title

Deploying Agents Without Writing Code: the Deepagents CLI

Goal

Show the student the alternative path: a "no Python" deployment that uses config files (deepagents.toml, AGENTS.md, skills/, mcp.json) instead of code. Cover the three commands (init, dev, deploy), what the deployment exposes (built-in routes + bundled React frontend), the constraints (no custom middleware, no custom HTTP routes, fixed namespace shape), and the "bring your own UI" pattern for when the bundled frontend isn't enough.

How to run this lesson

• Frame as a contrast with module 1: the code path is flexible but verbose; the CLI is convention-based and fast. Both produce a LangSmith deployment behind the scenes.
• m6.1 (concepts): walk the student through deepagents.toml sections, the files (AGENTS.md, mcp.json, skills/, subagents/, user/), and the three commands. Mention that under the hood the CLI generates a langgraph.json + handler code from the template.
• m6.2 (exercise): hands-on with dacli_tutor. Install CLI, set up .env, run deepagents dev, edit a skill, deploy. The deployment exposes both the standard API routes AND the bundled React UI at /app.
• m6.3 (bring your own UI): use deep_tutor's UI locally pointed at the CLI deployment. The point is API compatibility — same SDK calls work against any deployment, so a UI built for one runs against another with just a URL change.
• Hammer the contrast: code path = Python code + langgraph.json; CLI = config files only, no Python. Same underlying deployment server.

Key concepts to cover

• deepagents.toml — the config file that replaces langgraph.json for CLI deploys
• AGENTS.md — agent persona, loaded as system prompt
• skills/ — progressive-disclosure knowledge directories
• mcp.json — MCP tool servers; CLI supports HTTP/SSE only (no stdio)
• subagents/ — optional sub-agent definitions
• user/ — per-user memory directory; populated by the CLI's user-memory scoping
• The three commands: deepagents init (scaffold), deepagents dev (local), deepagents deploy (production)
• What the CLI deployment exposes: standard Agent Server API + bundled React UI at <url>/app
• [auth] provider options — anonymous, supabase, clerk; per-user identity comes from runtime.server_info.user.identity

10. The per-user namespace shape — (assistant_id, user_identity); auth identity IS the namespace

11. CLI constraints — no custom middleware, no custom HTTP routes, no Python code; trade flexibility for speed

12. The "bring your own UI" pattern — local UI server uses the SDK to talk to the CLI deployment

13. The seed file (_seed.json) — skills + AGENTS.md are bundled at startup; no hot reload during deepagents dev

14. Hub seeding — LangSmith Hub repo is seeded once per first deploy; refreshes require deleting the hub or editing via Hub UI

Tone guidance

Contrast-driven. Students just spent five modules on the code path. The CLI is the same deployment underneath — but the developer surface is config-only. When the student asks "but how do I do X?", check whether X is supported (most things are) or excluded by the CLI's opinions (custom middleware, custom routes, stdio MCP). If excluded, the answer is "drop to the code path."

Reference material

Full reference material is in information.md in this directory. Read it before answering factual questions.