AI Agent Engineer Intern (Claude, MCP, Evals)

Compensation: No equity

Unpaid internship · 3 months · 5–10 hrs/week · Remote (US) · Academic credit available

Zapp Studios AI Agent Lab — Fall 2026 Cohort

Zapp Studios builds growth marketing and software as one system for small businesses — restaurants, mobile pet care, e-commerce, home services. Real clients, real revenue, one small AI-native team.

The AI Agent Lab is a 3-month, part-time program where you learn to build with Claude — Claude Code, the Anthropic API, MCP servers, agent evals — and then point those skills at an actual business problem for an actual client. You finish with agents you shipped, a portfolio you can show, and working command of tooling most people are still just reading threads about.

This is unpaid, and we'll be straight with you about it. What you get is the training, real client work, the portfolio, and a written reference from someone who watched you build. If your school gives credit for an internship, we'll do the paperwork.

How the three months run:

• Weeks 1–3 — Learn. Guided ramp on Claude Code, context and prompt design, tool use, MCP, and evals. Structured, with checkpoints.
• Weeks 4–8 — Build. You take a scoped, real problem from a Zapp client or product and build the thing that solves it.
• Weeks 9–12 — Ship. Deploy it, measure it, demo it. It goes in your portfolio.

In this track you'll:

• Build agents that do real work — one that reads a supplier invoice and files every line against a real ingredient, one that drafts and QAs client campaign copy, one that watches a booking funnel and flags what broke
• Write the tool definitions, MCP servers, and context plumbing that make an agent reliable instead of a demo
• Build evals, because "it worked when I tried it" is not shipping
• Take one agent all the way into a live client workflow and watch a real person use it

You probably:

• Write Python or TypeScript and can read an API doc without hand-holding
• Have built something with an LLM already — a bot, a script, a hack — and have opinions about why it was flaky
• Have high agency and are comfortable when the problem is vague
• Care that the thing works for the business, not that it's clever

Nice to have: MCP, retrieval/RAG, eval harnesses, Claude Code as a daily driver, or having shipped anything to real users.

Everyone in the Lab gets: weekly working sessions with the founder, one real shipped project, and a written reference. A Claude subscription is not included: from Week 2 you'll need your own, the $20 plan to start, and be prepared to upgrade later in the program.

5–10 hours a week, remote, async-friendly, on your own schedule. Cohort runs Sep 14 – Dec 11, 2026. Applications close Sep 7.

To apply: in your application note, do two things — (1) link the single most impressive thing you've built, grown, designed, or sold, with one paragraph on what it does and what was hard about it; (2) confirm that unpaid, 5–10 hrs/week, Sep 14 – Dec 11 works for your schedule. Applications without a link get read last.