Multi-agent systems. The orchestration that ships.
Moonlabs is the operator-led AI Academy in Derby. We run three live companies — Homemove, home.co.uk and homedata.co.uk — and we teach twelve students per cohort to ship a real AI product, sell it to a real customer, and raise on it. Three pillars: Coding, Commercials, Investment. Twelve weeks. £6,000.
Moonlabs is what we are. Two operators — James Freestone and Louis O’Connell-Bristow — who run Homemove, home.co.uk and homedata.co.uk. We run multi-agent supervisor / worker flows in production at all three businesses — the failure-mode design on this page is what stopped our own systems from spiralling when we first put them in front of real users.
The Academy is what we do. A twelve-week, in-person, twelve-student cohort in Derby. You build a real AI product. You sign a paid pilot on it. You write a deck and a financial model. You leave with a deployed system, a paying customer reference and a live investor pipeline. Coding, Commercials, Investment — the three pillars taught in equal weight every week.
Why this page exists. Single-agent systems hit a ceiling past two or three tools, past a long task horizon, past the point where context becomes the bottleneck. Multi-agent is fashionable to talk about and rare to ship well — almost every multi-agent demo on Twitter is one prompt away from collapsing. You leave the Academy with a multi-agent system in production that survives a real production day — or as the founder of a multi-agent product before most of the market figures out the orchestration.
Coding · multi-agent at production fluency
Supervisor / worker patterns with deterministic supervisors and scoped worker context. Agent-to-agent protocols. Shared memory and state. Cost-aware orchestration (budget caps, recursion limits, model tiering, message-rate ceilings). Trajectory-level evals (replay-based, intermediate-state checking, end-to-end success metrics). LangGraph, AutoGen, CrewAI, provider-native agent SDKs — whichever fits the project. A deployed multi-agent system by week twelve.
Commercials · multi-agent as a productised offer
Multi-agent “digital workers” that replace a small ops team are the highest-ticket AI sale right now — SMBs pay £30-100k/year for one that genuinely works. Pricing per workflow replaced, the discovery call, a one-page pilot, the first paying customer. A paid pilot by week six — for many graduates, sold as a productised “your-back-office, run by agents” offer.
Investment · raising on multi-agent products
Sierra ($4.5bn, customer support multi-agent), Adept ($1bn before Amazon acquihire), Cognition / Devin (multi-agent SWE), /dev/agents (founded by Android co-creator on the multi-agent thesis), Crew AI, Lindy ($50m), Magnetic-One from Microsoft, 11x ($50m, multi-agent SDR) — multi-agent is the loudest live thesis in venture. Cap table, ten-slide deck, financial model. A live investor pipeline by demo day.
Common questions.
When is a multi-agent system actually the right answer?
Not as often as the framework vendors suggest. The honest test: does your task have natural sub-roles, a horizon longer than one prompt, and benefit from explicit memory across steps? If yes, multi-agent. If not, a single agent with tools is almost always cheaper and faster to ship.
Which frameworks do you cover?
LangGraph and AutoGen as the production-default options; CrewAI where it earns its place; provider-native agent SDKs (OpenAI Agents, Anthropic agents) where the orchestration fits inside them. The course is provider-agnostic on framework choice.
How does this fit with the wider agents and orchestration curriculum?
Multi-agent is the deepest end of the agents curriculum. The agents course page is the general framing; the LangGraph page covers the orchestration framework specifically; this page focuses on the multi-agent design question.
Can multi-agent systems be debugged at all?
Yes, with discipline. The observability requirements are higher than for single-agent systems — structured traces with agent role tags, intermediate-state snapshots, replay tooling. We teach the patterns that make a stuck flow debuggable in minutes.
Will the frameworks I learn here go out of date?
The tooling will. The disciplines — role separation, evals on the trajectory, cost discipline, failure-mode design — will not. We refresh the tooling each cohort.
More Academy entry points.
The Academy is one course with many doors. Each of these pages is a different entry point into the same twelve weeks.
Build it. Sell it. Raise on it. In twelve weeks.
Tell us what multi-agent system you would build and what would go wrong first. James and Louis read every application personally and reply inside the week.
© 2026 Moonlabs Incubator. All rights reserved.