1 · The Big Idea — Why a Multi-Agent AI Operating System

Part 1 of 6 · ~6 min read

If you’ve used ChatGPT, Claude, or Gemini, you’ve experienced AI as a conversation. You ask, it answers. The relationship is one human, one model, one chat window.

That’s a useful pattern. It’s also a deeply limiting one — like trying to run a business by hiring a single very smart consultant who can only work on one thing at a time, can’t remember what they did last Tuesday, and refuses to collaborate with anyone else.

What I built in April 2026 is the opposite of that pattern. It’s a small operating system for a team of AI workers. Different models for different jobs. Tasks that span hours. Workflows that chain together like a relay race. Memory that persists across sessions. The whole thing runs on a few top-tier AI subscriptions, a couple of personal computers, and trivial cloud spend — total cost less than hiring a single junior analyst.

This article explains the idea behind it. The next five articles get into how it works, what apps run on it, and what it took to build.

The shift you’ve already noticed

Watch how a serious AI user actually works in 2026.

They open Claude in one tab to draft an email. They open ChatGPT in another because it’s better at math. They have Gemini available for things that need fresh web data. And they keep a CLI agent (Cursor, Codex, Claude Code) running in their terminal for code changes.

That’s already four AI workers. They’re just not coordinated. The human is the integration layer, copy-pasting between windows, holding context in their own head, deciding which tool to ask.

The next obvious step is to make those AI workers talk to each other directly. Not through a human’s clipboard. Through software.

That’s what an AI orchestration platform does. It’s the layer that lets you say “research this question across three models, summarize the findings, generate a report, and email it to me by 9 AM” — and have all the pieces happen automatically, in the right order, on whatever machines have spare capacity.

That’s the big idea. Single AI tool → coordinated AI team. Conversation → choreography.

What “multi-agent” actually means

The word agent gets thrown around a lot. Here’s the operational definition I find useful:

An agent is anything that can take a request and produce a result — without you watching every step.

By that definition, ChatGPT is an agent. So is a Claude Code session that edits five files. So is a search engine like Google’s AI mode. Each one has its own strengths. None can do everything.

A multi-agent system is a coordinator that knows about a fleet of these agents and can pick the right one for each job, route work between them, remember what each produced, and present the combined result.

That’s it. The coordinator is the operating system. The agents are the apps. The user types one request and the system figures out which apps to run.

Why this matters now (specifically, in 2026)

Three things lined up in the past 18 months that made this possible for one person to build:

Models reached a usable cliff. Claude Sonnet 4 and Gemini 2.5 Pro crossed a threshold where they can complete medium-complex tasks end-to-end without supervision. Before that, every task needed a human in the loop. Now, many don’t.
Code-writing agents became reliable enough to build infrastructure with. Tools like Claude Code can not just write a small backend service but operate it — refactor, debug, deploy. That collapses the “I have an idea but don’t have engineering capacity” gap.
Edge compute got cheap and globally available. Cloudflare Workers, Workers AI, D1, R2, KV, Queues — a single individual can stand up a distributed system that scales globally with no servers, no DevOps team, and a monthly bill in the low single digits.

Stack those three and a small team — or one person, with an AI partner — can now build infrastructure that would have required a Series A round two years ago.

ResearchOS is what I built when I noticed all three of those shifts and asked: what’s the smallest useful thing that demonstrates this?

The shape of what it does

Without going into how, here’s what the platform does once it’s running:

You can type a research question — “What are the investment opportunities in the AI industry chain?” — and it will quietly spin up dozens of AI workers, each researching a sub-question, each generating sources and analysis, all stitched together into a navigable tree of findings.
You can describe an investment-research workflow — “every morning, my ESG analyst should produce a carbon-policy update, my derivatives strategist should track basis trades, the chief economist should write a coordination memo” — and the system runs that workflow automatically on a schedule, picking the right model for each role.
You can capture an idea on a kanban — “investigate stablecoin regulation in Singapore” — and turn it into a full research session with one click. The result lands in your dashboard the next time you open the page.
All the work happens across a small fleet of machines you control — including ordinary laptops and Mac Studios you have lying around. The system load-balances across them, retries failed tasks, recovers from crashes, and keeps the whole thing running without supervision.

None of this requires you to write code. Once the platform exists, you operate it the way you’d operate a Notion workspace or a spreadsheet — by describing what you want.

Two real apps already running on it

The platform has two live applications, both of which I built and use myself:

Infinite Research — type a question, get a recursively explored tree of answers, all with sources. Detailed in Article 3.
AI Institute — a multi-agent investment research operation. A team of role-played AI analysts (compliance officer, ESG analyst, derivatives strategist, chief economist, real-estate analyst, theme researcher) produces daily morning briefings, weekly reports, and ad-hoc investigations. Detailed in Article 4.

Both apps share the same underlying platform. Both are usable today. Both were built in days, not months, by one person with an AI partner.

What I want you to take away

If you read nothing else in this series, take these three things:

The single-AI-chat era is ending. Coordinated teams of AI workers, running on cheap infrastructure, operated through plain-English workflows, are becoming feasible for individuals — not just well-funded companies.
The cost has collapsed. What I built — across 10 days, including two production applications — runs on three top-tier AI subscriptions (Claude, ChatGPT, Gemini, each ~$200/month), a couple of Mac Studios and a MacBook I already owned, near-zero Cloudflare spend, and a few third-party APIs. The dominant cost is the AI subscriptions, not the infrastructure. A junior human analyst would cost an order of magnitude more.
The interface is becoming the operating system. You don’t write code on this kind of platform. You describe what you want, and a layer of AI agents figures out how to do it. Whoever masters this interface over the next 24 months will have outsized leverage.

If any of that lands, the rest of this series will show you how it works, what it builds, and what it took to put together.

Next: How It Works — A Restaurant for AI