Assistant to AI Agents

Designed Make’s first AI Assistant, then AI Agents, and brought agent setup back into the canvas.

3 min read

The Make.com AI Agent beside its reasoning steps: input, thinking, and the tools it called

When I started in early 2024, Make had no AI assistant. There was no playbook, the industry was still learning, and we were falling a little behind. I was the designer next to about eight developers and a PM, and over two years I designed the foundation for two things: the AI Assistant, which built scenarios for you, and AI Agents, which work inside them. The biggest lesson came from a mistake. For a while, we took agent setup out of the canvas.

Client
Make
Role
Design Lead
Team
8 Devs, 1 PM
Timeline
2024 – 2025

01

Two kinds of AI

The Assistant was a layer above the canvas. You asked it in a chat, and it built the whole scenario or answered your question, but it had no control over what happened while the scenario ran.

An Agent is a module on the canvas, working directly inside the scenario and doing the work during the run. Every other module on the canvas was deterministic. Agents are not.

02

The Assistant

Surprisingly, a lot of people used the Assistant, mostly for simple scenarios. It made mistakes, and we saw them in support tickets and in our own testing. It only built deterministic scenarios, its success rate was low, and it hallucinated. People couldn’t see how it built anything, so they didn’t trust it with workflows that mattered. We added a rating on every answer, a global undo, and a copy of the scenario saved before the AI changed anything.

After a few months many people stopped using it. The trust wasn’t there, and it was slow.

The Make.com AI Assistant chat building a scenario from a request
A hexagon AI Agent module on the Make.com canvas, connected to its tools

03

A page outside the canvas

The first Agents, at the end of 2024, were set up on their own page. You’d create an agent once as a global agent, reuse it, and give it local instructions in each scenario. The setup was smart and worked correctly. People didn’t understand it.

All building in Make happened in the canvas. Taking part of it somewhere else was a mistake.

The first Agents page, outside the canvas: an agent’s system prompt, context, MCP, tools and variables beside its test chat

04

30+ interviews

I was in more than 30 user interviews on this project and ran about five of them. Mostly I took notes and asked a lot of questions: my goal wasn’t to run them, it was to gather as much feedback as I could. Some were very tough to hear. People said the agent felt like a black box. They didn’t trust what it was or what it would do, so they wouldn’t give it complicated automations or sensitive data like HR or finance.

The main decision that came out of them: agent setup lives in the canvas only.

05

Back into the canvas

Moving everything from the setup page into the canvas was the hard part. Setup couldn’t become brutally complicated, and it had to show what was actually happening and when.

Shape
Every other module is deterministic and agents are not, so agents got their own shape: a hexagon. People just had to learn the difference.
Test chat
You can chat with an agent and train it in a separate chat before you put it into the canvas. That chat was how we tried to build trust.
Reasoning
A reasoning view shows, step by step, what the agent did and why.

The canvas UI took more effort than anyone will notice. The information is dense, and we wanted to give people the freedom to customise an agent while keeping it for everyone, not only for technical people.

Agent setup inside the canvas: connection, model, prompt, mapped user message and tools
The reasoning view: step by step, what the agent did and which tools it called

06

Where it went

Each version left the one before far behind, in adoption and in success rate. The numbers stay private.

For a while the Assistant and Agents ran side by side. We wanted to switch the Assistant off, because we knew it would grow into MAIA: a smart assistant above everything, living in the canvas for now, with plans to expand. MAIA was built on what we learned with the Assistant. I helped with some visual work, like its mascot, and gave my opinions; the credit for MAIA belongs to its designer.

The Make.com AI Agent with the apps it works with: Make, Airtable, Make Code and Slack

07

Looking back

At the end of 2025 I moved to the design system full time and handed Agents to a new designer. Today’s Agents still stand on the foundation I set, with setup in the canvas, and the flow and UI are much more polished. After all the hiccups, with no playbook from the industry, agents are now smart and powerful in automations. I’m proud of how that turned out.

If I started again, I’d put agent setup in the canvas from the start, not on a separate page. That’s where the interviews led us anyway.

It was also the first project in my career where I worked closely with AI myself, beyond GPT for copy or translations. Vibecoding for designers was far behind where it is now, and we used v0 by Vercel to prototype UI ideas fast.

Next case studies

Interested in working together? I’d love to hear about your project or team.