Most 'multi-agent' AI systems aren't multi-agent

Most products that call themselves "multi-agent AI" aren't. They just sound better that way.

What they usually are: a series of steps, with a single AI doing one specific job somewhere inside. There's nothing wrong with that — most problems don't need anything fancier. But the label matters, because it quietly changes what a buyer, a recruiter, or a hiring manager assumes you built.

I built one of these tools — Sonar, which finds sales leads — and I had to talk myself out of calling it multi-agent. This is the explanation I owed myself.

What the words actually mean

People sling four terms around as if they're interchangeable. They're not. From simplest to fanciest:

• An agent is a single AI working in a loop, deciding its own next move based on what just happened. The key part: it decides what to do next.
• Multi-agent is several of those running at once and coordinating — usually a "manager" AI handing tasks to "worker" AIs that report back. More than one independent decision-maker, talking to each other.
• An agentic workflow is a fixed set of steps you wrote, with an AI plugged in at certain points. The AI does its bit but doesn't choose the order. This is what most real products actually are.
• A multi-stage pipeline is an assembly line: fixed steps, with an AI called in to do a specific job at some of them (sort this, write that). No loop, no AI picking the path. This is what most "AI agent" products really are — mine included.

Anthropic's engineers have been blunt about this: the main thing a true multi-agent system buys you is the ability to throw lots of parallel work — and lots of money — at a problem. If your task fits in one AI's head, or the work doesn't split into independent chunks, going multi-agent just makes it more expensive.

Why everyone overclaims

Because the bigger word sells. "Multi-agent" and "AI agents" are hot on LinkedIn, in pitch decks, in job posts. The phrase signals sophistication to people who can't tell the difference — and most people, including plenty of recruiters and investors, can't.

So one AI call with a tool ships as "agentic." Two in a row ship as "multi-agent." Five ordinary scripts plus one AI and an approval step becomes an "autonomous multi-agent intelligence platform" on the landing page. The honest end of the field has been pushing back on this for over a year. The marketing is still winning.

My tool, honestly

Here's what Sonar actually does, start to finish: it gathers leads from a handful of public sources, has one AI sort the promising ones, finds a contact email, has a second AI draft a personalized opener, sends that to me to approve or edit, then sends the message.

From the outside, that looks multi-agent — six steps, two different AI jobs. It isn't:

• No AI talks to another AI. Each step hands its finished work to the next, like stations on a line.
• Nothing "decides" what runs next. The order is fixed — by me, in plain code.
• Each AI call does one small job and stops. It doesn't ponder, doesn't wander off on its own.

What Sonar really is: an assembly line with an AI at two of the stations, plus a human (me) approving before anything goes out. That's a mouthful, but it's the truth, and every word in it is doing a job.

When multi-agent is the right call

There's a real case for it. Anthropic's rule of thumb is the cleanest one I know: reach for multi-agent when the task breaks into lots of independent pieces, there's more to chew on than a single AI can hold at once, and you can afford the bigger bill.

"Search two hundred sources at the same time, then combine the findings" — that's a genuine multi-agent job. "Sort a thousand reviews, then write an email for each" — that's not. It's an assembly line, and forcing it to be multi-agent just adds cost and things to break.

Most real products are the second kind. Most of them should stay that way.

Why I built mine the boring way

• Cheaper. One AI call per lead, no "manager" AI to pay for.
• Easier to fix. When a draft comes out wrong, I look at one step — not a tangle of AIs talking to each other.
• Predictable. I know what a thousand-lead run costs because I know exactly how many AI calls happen. A multi-agent version could quietly balloon.
• Honest in an interview. When someone asks me to walk through how it works, I can. It fits on a whiteboard.

The boring choice wasn't a compromise. It was the right tool.

The takeaway

Most things that claim to be multi-agent are assembly lines with extra steps. And most assembly lines that genuinely would need to be agents would be better off staying simple.

Read how it actually works, not the headline.