No Robots, No Sparkles: How to Draw AI Agents

A brief history of AI (drawing)

I often draw multi-agent systems as part of my product work and in the UX for AI cohort. If you've been reading my posts over the last 2 years, you've seen me draw agents in all kinds of ways:

But recently, multi-agent systems really became the focus of my work, so I had to get consistent in how I draw the various agent types and the humans who interact with them. My requirements were simple:

• Stupid-easy to draw. A marker, a whiteboard, 5–10 seconds.

• Instantly different from a human. Even next to a stick-figure face, nobody should confuse the two.

• Clear about the job. You should grok the agent's function instantly, without the label.

Unfortunately, our field has not produced anything that remotely matches these requirements (or even anything consistent, for that matter... and I am just as guilty of this -- see the exhibit above).

So I did what every self-respecting forward-deployed product manager does.

I asked Claude.

"So... how would you want to be represented as an AI agent? Robot? Sparkles? Something else?"

No robots, no sparkles

The answer came back fast, and it was firm:

Then it pointed back at my own sketches. I had drawn each agent as a robot-shaped head with a symbol: an S for the Supervisor, a magnifying glass for the Worker. (The Supervisor was my "Vision" robot, more complicated than the others and more complicated than it needed to be. A pain to draw. Every. Single. Time.) Claude's verdict: "Your own sketch already has the right idea. It tells the viewer what the agent does, not what it is. A planner gets a checklist, a judge a gavel, an executor a lightning bolt. Function over face."

My agents were trying too hard to be robots.

(Yeah, yeah, "the chatbot is trained to deny being a character." Whether an AI can "want" anything other than to make more paperclips is beside the point. We're trying to come up with a useful joint AI-human presentation convention here. Get with the times already!)

The finger pointing at the moon

Claude proposed a rounded tile with a job symbol inside and a small port on each side, like a component that plugs into a pipeline. No face, no eyes:

I looked at it and thought: erghh... those ports are ears. It still looks a bit like a robot head.

And I decided that's fine. Better than fine. The ears are just anthropomorphic enough to make sense, stupid-easy to draw, and immediately distinguishable from a human, even a stick-figure face. And the symbol tells you the job. It's just like what I was drawing, but with no eyes, so the symbol is bigger and easier to draw.

What matters with agents is one thing: what's in their head. A checklist, a lightning bolt, a magnifying glass, scales, or a gavel. That reflects where enterprise agents really are today: they are defined by a job. They are created to be useful, not to be a TV personality.

The icon is the finger pointing at the moon, not the moon itself. It points at what the agent does. Not a robot with a body. Not magic sparkles.

The convention

The ears got a job too: they act as terminals, like in a flow diagram. Flow goes left to right: the left ear is an input, the right ear is an output, and an agent gets as many as it needs, within reason. For example, a Judge gets three: one in, two out.

Important: ears mark kinds of flow, not message counts. It's all about the key steps in the flow. So don't make agents with 10 ears. Two or three? Fine. More than that, split it into two agents. (10 ears would just be weird. You'd hear like everything. Like the CIA maybe. Or your smart speaker.)

Ahem. Right. Moving on...

The resulting diagram is simple to sketch, and an odd shape only shows up where it means something, like the Judge's escalation path.

In practice: Mobot at Sumo Logic

Here is the convention applied to a real system: Mobot, the multi-agent investigation system I designed and built at Sumo Logic. It cut mean time to resolution from 60 minutes to under 3 and false positives by 90%. The agents don't call a human unless they absolutely have to:

It's not human-in-the-loop, but a truly automated SOC investigation with a human on the loop. I break down that difference, step by step, in Human-in-the-Loop in Agentic Investigations: A Snowball Sprint Storyboard Analysis.

That is the point of drawing it this way: the tiles do the work, and the face makes the call when the agents can't.

In one ear, out the other

1. Every agent gets the same tile. A rounded box with ears. Same shape means same kind of thing.

2. The symbol inside is the job. Checklist for the Planner, lightning bolt for the Executor, magnifying glass for the Analyst, gavel for the Judge. (Resist adding eyes. Or a mouth. Control! You must learn control!)

3. Left ear in, right ear out. One of each by default.

4. Add an "ear" only for a different kind of flow. The Analyst takes a plan and data. The Judge sends work on to act, or escalates it to a person. A Planner creates ten steps but still has one output ear, because all ten are the same kind of flow.

5. Only humans get faces. The human face is, ahem, "the face" of the flow. It's where accountability sits, and the reader should find it at a glance.

Why this matters

Over the last 2 years, I realized something important: how we draw agents determines how people think about them. Draw a robot, and viewers read intentions into it; they trust it or blame it like a person. Draw sparkles, and they stop asking how "The AI Magic" works. Both are clear design failures, because robots and sparkles hide the most important questions in any agentic system: who does what, what gets handed off, and where does a human decide?

A flow diagram made up of agent tiles with job symbols helps us answer those questions. It doesn't flatter the agent or imbue it with magical powers. It simply shows which part of the system does which job, and who owns the outcome.

You can draw the whole flow in under a minute. (And who knows, it might just save your team a few months and a few million bucks of going down the wrong path.)

I've now added this convention, proudly, to my LinkedIn banner and to the curriculum of the UX for AI Professional Certification.

Draw the job, not the robot.

Our new UX for AI Advanced Course on Designing Agentic Systems starts in November. Head over to our Certification page to join the waitlist and get first dibs and the lowest price. (I'm not telling you to hurry or anything, but our first UX for AI cohort sold out in 11 days. Just saying.)