Internship

AI Agent Orchestration Internship

We no longer teach coding. In 10 weeks you learn to direct AI agents that ship real software and find its buyers, from Byblos, Lebanon.

5
Years running
13
Seats a year, at most
10
Weeks
0
CVs read

Definition

What Is AI Agent Orchestration?

Directing several AI agents as one team: splitting the work, giving each the right context and tools, checking the output and shipping the result.

Curriculum

Ten Weeks, Spec to Customer

Five stages, start to finish. Every week ends with shipped, reviewed work.

1. Specify · Weeks 1 to 2

  • Week 1 · How agents really work. Context, tokens, tool calls, why agents fail. You ship: a real bug fixed by one agent, every diff reviewed.
  • Week 2 · Specs and context. Acceptance criteria, repo memory, reusable skills. You ship: a repo an agent can work in cold.

2. Delegate · Weeks 3 to 4

  • Week 3 · Tools and permissions. MCP servers, sandboxes, secrets, prompt injection. You ship: a tool server with least-privilege access.
  • Week 4 · Agent teams. Planner, workers, reviewer, parallel runs in git worktrees. You ship: one feature built by three or more agents.

3. Verify · Weeks 5 to 6

  • Week 5 · Evals and proof. Tests as gates, browser checks, adversarial review, traces. You ship: an eval suite that catches a seeded bug.
  • Week 6 · Cost and recovery. Budgets, model routing, caching, checkpoints. You ship: a run that is killed midway, resumes and reports its cost.

4. Ship · Weeks 7 to 8

  • Week 7 · A real client brief. Human sign-off, security review, deployment and handover. You ship: real work for a MoonDev client.
  • Week 8 · Live factory run. An unseen brief, run live in front of the team. You ship: your factory, its post-mortem and a lesson you teach back.

5. Distribute · Weeks 9 to 10

  • Week 9 · Find the buyer. Positioning, buyer profiles, research agents that cite every fact. You ship: a verified target list for a real MoonDev offer, built with our sales team.
  • Week 10 · Reach them. A person approves and sends every message, and you fix the agents from real replies. You ship: a demo day judged on meetings booked, not messages sent.

Partners

Universities and Partners

Institutions we have worked with.

The challenge

Build the Factory, Not the Product

Anyone can ship an app with AI now. So build the system that ships apps, then run it live on an unseen brief.

The brief

Build a factory: a system of AI agents that turns a one-page client brief into a deployed web app, with nobody editing the output by hand. The app is the evidence. The factory is the submission.

  • Write a one-page brief for something your assigned client type would really pay for, then run your factory on it.
  • Use at least three distinct agent roles, for example planner, builder and reviewer. The reviewer must be able to reject work.
  • Once the run starts you may answer questions the factory asks you, but you may not touch its output. Every intervention is logged.
  • The factory records what each agent did, what it cost and what failed.
  • From the same brief, the factory also produces a launch kit: who the buyer is, 20 real target companies with a source for each, and a tailored first message for each. Submit the drafts. Do not send them.
  • Any model, any tools, any stack. The result must be a live URL.

Your unique assignment

Register below and you receive a client type and one hard constraint by email. A constraint could be that the run stops itself before passing a budget you declare, or that it resumes after being killed midway.

Your 7-day countdown starts when you register, and that timestamp is part of the evaluation. Do not share your variation: submissions are fingerprinted, and identical or near-identical systems are disqualified.

What to submit

Email info@moondev.solutions with the subject “[Your Name] - MoonDev 2026 - Agent Challenge” and include:

  • The repository of your factory
  • The live URL of the app it shipped from your brief
  • The full run log, with cost and every human intervention
  • The launch kit: buyer profile, 20 sourced companies and unsent first messages
  • A 300 to 500 word post-mortem: what your agents got wrong, what you caught, and what you would refuse to automate and why
  • Evaluator credentials, if the app has a login

A late submission is still reviewed, but on-time delivery counts. If something serious comes up, email us before the deadline.

How it is scored

  1. Autonomy. How far the factory gets without you.
  2. Verification. How it catches its own mistakes before you do.
  3. Recovery. What it does when a step fails or the brief changes.
  4. Cost. What each shipped feature cost.
  5. Product judgment. Whether a real client would accept the result.
  6. Distribution. Whether the launch kit names real buyers and says something they would answer.
  7. Understanding. Whether you can explain what your system did and why.

A bigger app scores nothing on its own.

The live run

The live run replaces the interview. If you are shortlisted, you share your screen for 60 minutes and we hand you a brief you have never seen. Partway through, the client changes their mind. Your factory has to absorb the change and ship to a live URL.

We then spend 15 minutes on what happened. We open a file or a failed step and you explain it. A borrowed system does not survive this part.

Admission, step by step

  1. Assignment. Enter your email and receive your client type and constraint.
  2. Build. 7 days to build your factory and submit.
  3. Review. MoonDev engineers read every submission, and everyone who submits gets feedback.
  4. Live run. Shortlisted candidates run their factory on an unseen brief, then explain it.
  5. Seat. The strongest runs get the seats, at most 13 a year.

No CV, no grades and no degree requirement at any step.

Why distribution is in the program

Writing code is now cheap, so software nobody hears about is worth very little. For the last two weeks you work with MoonDev's sales team on a real MoonDev offer.

Your agents research companies and draft messages. Our sales team checks the facts, rewrites what is weak and sends every message. Volume is capped on purpose, and demo day is judged on meetings booked, not messages sent.

For universities

Universities can propose students for the program. Every student takes the same challenge and no seats are reserved, which is what makes a seat worth something on a graduate's record. Supervision and credit paperwork are agreed before the cohort starts. Talk to us about your students.

Why we ended the coding internship

For 5 years we ran this internship to teach people to build software. AI now writes routine code faster than an intern can learn to, so a coding internship no longer prepares anyone for this industry.

The scarce skills are deciding what to build, directing agents, proving the result is right and getting it to buyers. The same skills sit behind our AI and automation work.

What we did with these partners

  • Lebanese University: Academic and student initiatives.
  • USEK (Holy Spirit University of Kaslik): University partner.
  • NDU (Notre Dame University, Louaize): University partner.
  • LIU (Lebanese International University): University partner.

Apply

Five Steps, No CV

Assignment · 7-day build · Review · Live run · Seat

Enter your email to receive your unique assignment. Your 7-day countdown starts on submit.

FAQ

Internship questions

Who can apply for the AI agent orchestration internship in Lebanon?

Students, recent graduates and self-taught builders who can read code. There is no degree requirement and we do not read CVs. You are judged only on the agent system you build and how it performs in the live run.

Why did MoonDev stop running a coding internship?

Because AI now writes routine code faster than an intern can learn to. Teaching that skill alone no longer prepares anyone for this industry. What stays scarce is deciding what to build, directing agents, proving the result is right and getting it to buyers, so that is what we teach.

Do I still need to know how to code?

You need to read code, not type it quickly. An orchestrator has to spot when an agent is confidently wrong, and you cannot review what you cannot read. In the live run we open a file your system wrote and ask you to explain it.

Why are there only 13 seats a year?

It is limited on purpose. Every intern's work is reviewed by MoonDev engineers each week, and that does not scale. We take 13 people at most, and we leave seats empty rather than lower the bar.

Can my university send students to the internship?

Yes. Universities can propose students, and supervision and credit paperwork are agreed with the university before the cohort starts. Every student takes the same challenge, and no seats are reserved for any institution.

Do I have to pay for AI tools to apply?

No. Any model and any tool is allowed, including free tiers and models you run locally. Cost is one of the scoring criteria, so a more expensive run does not score higher.

Is the internship paid?

Compensation and start date are shared with the offer, because they depend on the cohort and the project you join. If you need this before starting the challenge, email info@moondev.solutions and we answer directly.

Can I do the internship remotely?

The challenge and the live run are fully remote. The MoonDev team works from Byblos, Lebanon, and whether your internship is on-site, hybrid or remote is agreed after the review.

Thirteen Seats. One Way In.

Your 7 days start when you register.

Take the challengeUniversities: talk to us

info@moondev.solutions+961 81 81 00 57WhatsApp