Beyond the Chat Box: What Building a Multi-Agent Studio Taught Me About Human-Centered AI

Early on in building an AI-native comic studio, I set up what felt like a natural experiment: a swarm of agents inside a single chat. I could talk to the lead agent, and the others would think out loud as they collaborated. Elegant in theory. In practice, I drowned. Ideas, drafts, critiques, and tangents poured in faster than I could read, let alone respond to. I was either overwhelmed or, when I tried to keep up, the bottleneck slowing everything down.

Neither is what good collaboration looks like.

I want to share some lessons from this work, because I think they generalize past the comic studio. If we’re going to build great AI-native experiences, we need to design well past the chat box. The question isn’t just “what does the user type and what does the AI say back?” It’s: how do we reduce cognitive load on people, amplify their impact, and shape how agents collaborate with humans and with each other?

A bit of grounding

The studio is a multi-agent environment for creating comics: characters, scripts, panels, layouts, dialogue, art direction, and the connective tissue between them. There are many specialized agents, organized into teams, working on shared documents and whiteboards alongside me. Even a single agent can produce ideas, feedback, and visuals faster than I can absorb. Multiply that by a roster of agents, and you have a UX problem hiding underneath an AI capability story.

The lessons below are the design dimensions I keep returning to.

Roles: what each agent brings

The first design move isn’t about interfaces. It’s about defining who each agent is. Their knowledge, their skills, their abilities, and what I’ve started calling their angle: the perspective they prioritize, the edge they bring. A character designer agent and a continuity editor agent might both review the same panel, but they’re looking for very different things. Without sharply defined roles, agents blur together and the human loses any sense of why a particular voice is in the room.

Roles also let us think about teams, and teams of teams. Different work calls for different configurations, and the workflow is rarely linear. There’s revisiting, jumping back, looping in someone new. Designing for that means accepting that the org chart is fluid.

Collaboration patterns, including agent-to-agent

Most discussion of AI UX focuses on the human-agent boundary. But in a multi-agent system, agent-to-agent patterns are part of the user experience too, because they shape what eventually surfaces to the human.

A few patterns I’ve found useful:

• Swarm discussion: a team of agents debates a problem before responding to the human, so what comes back is synthesized rather than a pile of raw takes.
• Consultation: an agent pulls in a specialist when needed, instead of pretending to be everything.
• Debrief on handoff: when an agent picks up work others have already touched, it actively asks the prior agents what happened, what was tried, what was rejected. Continuity without me having to be the human bus.

The thread tying these together: not every agent voice needs to reach the human. A boardroom doesn’t work when everyone talks at once, and neither does a chat.

Interaction surfaces: there’s more than chat

Chat is a default, not a destiny. Once you accept that not every interaction needs to be a typed exchange, the design space opens up.

Two examples from the studio. When I built a dedicated area for crafting reference characters, I reflexively dropped a chat into it. Then I noticed I was typing the same prompt over and over: “can you review and suggest improvements?” Why was I typing? A button would do it: one click, agent reviews, suggestions appear. The interaction was begging to be simpler.

Elsewhere, agents are most useful when they make a plan. I give them a goal; they propose an approach; I react to that. Sometimes they delegate across other agents or teams without me orchestrating it. Sometimes valuable interactions get tucked behind a button, a review step, or a draft surface, invisible until I want them.

The point: chat, buttons, dedicated screens, plans, delegation, review flows — these are all UX primitives for human-agent work. Choosing the right one is a design decision, not a default.

Shared artifacts: knowledge in the world

Don Norman’s distinction between knowledge in the head and knowledge in the world matters here. When humans and agents are both creating, the artifacts have to do real work; they’re how we coordinate without holding everything in our heads.

A few things that have helped:

• When agents add ideas to a document or whiteboard, they @-mention themselves. It’s instantly clear who contributed what.
• When an agent (or group of agents) is actively working on an artifact, an overlay signals it. No silent overwrites, no two editors stomping on each other.
• Every artifact has history and rollback. Ideas don’t get lost; experiments are safe.

The underlying principle: make the work visible, make authorship legible, and never make the human afraid to let agents touch something.

Tying it back to human-centered AI

We talk a lot about human-centered AI, and we should. But in a multi-agent world, the term needs to stretch. Reducing cognitive load on people and amplifying their impact depends not just on the human-agent interface, but on how agents collaborate behind it. The agent experience — how they hand off, debrief, consult, and divide work — directly shapes the human experience. They’re connected.

Roles, teams, collaboration patterns, interaction surfaces beyond chat, and shared artifacts: these are the dimensions I keep coming back to. None of them are solved. But naming them is the start of designing for them, instead of defaulting to a chat box and hoping for the best.