Build Your First AI Agent · Lesson 8 of 8

Lesson 8 — Ship it

60 minutes. Deploy your agent somewhere real. Watch it run. Iterate.

---

Where to deploy

Three deployment shapes, in order of complexity:

1. Your laptop, on a schedule. `cron` or Task Scheduler. Simplest possible. Good for testing.

2. A small VPS. $5/month at DigitalOcean / Linode / Hetzner. PM2 keeps the process alive. Good for serious production for a single business.

3. A container in your existing cloud. ECS Fargate / Google Cloud Run / Azure Container Apps. Pay-per-use. Good for scale or when you have other cloud infrastructure already.

For a first agent, option 2. Cheap, real, gives you ops experience without overkill.

The deployment checklist

Before shipping to production:

  • [ ] Run a 1-week shadow test (agent runs, drafts proposed, nothing actually sent)
  • [ ] Approval gate wired to a real review surface (Slack, email, admin UI)
  • [ ] Logging on every tool call (timestamp, input hash, output)
  • [ ] Error alerting (Slack ping when error rate > X)
  • [ ] Manual kill switch (a config flag that stops the agent immediately)
  • [ ] Budget alarm on Anthropic API spend
  • [ ] Backup plan if the agent goes down (humans handle the work manually)

The kill switch is the one most people skip. Make it. You'll need it within the first month.

What happens in the first week

The agent will make decisions you didn't predict. That's normal — it's a probabilistic system, not a deterministic one.

Three things to watch:

  • Approval rate. What % of agent drafts get approved as-is? Target 70%+. If lower, refine the prompt.
  • Time saved per review. How long does it take to review-and-approve vs. write-from-scratch? If it's not 3x faster, the agent isn't earning its keep.
  • Edge cases caught. Make a log of weird situations the agent handled badly. Each one becomes a refinement.
What to build next

Once one agent is humming:

  • Add a second tool to the same agent. Make it more capable.
  • OR build a second agent for a different workflow. Same pattern.

After three production agents, you understand the design space. After ten, you're operating at a different level.

You've finished

You can now:

  • Design an agent for a real business problem
  • Define tools the model can call
  • Build the tool-use loop
  • Wire it to email, calendar, datastores
  • Handle errors and recover from failures
  • Gate actions behind human approval
  • Ship the whole thing to production

That's a skill very few SMB operators have. Use it.

What's next

The MCP Practitioner Course — go deeper on Model Context Protocol. Build remote MCP servers your agents can connect to. Package and sell MCP services to SMBs.

Champion's Bootcamp — apply this to a real business problem at your company. Ship a working prototype. Pitch it to leadership.

→ Course complete.

Previous: Lesson 7 — Approval gates and human-in-the-loop · Back to the course