The Strategy and Patterns Shaping AI in Chrome
A conversation with Monica Gonzalez on agentic workflows, the compose plate, and AI as design's newest material
Monica and I go way back to our grad school days at Carnegie Mellon. She’s spent the last eight years on Chrome’s desktop team at Google, leading design and overseeing search and Gemini integration. When we sat down to talk about agentic workflows, what stayed with me was the strategic lens she brings — combining AI capabilities, human-centricity, and industry signals. We also got into some of the emerging patterns in this space: the “compose plate,” “voice of the agent,” and something we started calling AI liveliness, without ever quite landing on a real term for it.
Without further ado — enjoy!
From Industrial Design to Chrome
Kursat: Can you tell us about your journey as a designer — what brought you to where you are today?
Monica: I’m originally from Mexico, where I studied industrial design. My first job was designing trains, and I considered a master’s in automotive design before my boss suggested interaction design. I loved its focus on people and personal invention. After a master’s at CMU, I worked at Microsoft, Amazon, Trulia, and Smart Design. I’ve been at Google for eight years, managing the design team for Desktop Chrome, search integration, and Gemini integration.
Where AI Changed the Work — and Where It Hasn’t
Kursat: How has AI changed the way you actually work, day to day?
Monica: Personally, my role means I’m rarely as hands-on as I used to be, but with AI, I get to be more hands-on again. Before, making even a quick prototype would have taken time — you had to draw it in Figma and really noodle on it. Recently, I made a quick prototype to explain an idea: we wanted the Gemini side panel to work across tabs, so a thousand tabs could belong to different conversations. It was the easiest way to explain it. It didn’t look perfect, but the idea was there, and people could really see it. Almost everyone on that project ended up making their own prototype with different nuances, so it was interesting to experiment in real time with very little work.
At the production level, we still rely on Figma. Craft is something AI hasn’t fully conquered. I imagine it will eventually, but for now, we emphasize highly polished UI with tight rules for our design library. Chrome as a product seems simple, but it is very complex — editing often takes more time than creating. Chrome looks similar to how it did ten years ago, but its capabilities have changed. We try to keep it clean, familiar, and simple, because our job is to help the web succeed. The product recedes to the background.
Kursat: I can relate to that — AI is non-deterministic. It can just make up something that’s not part of the design.
Monica: Yes, it’s all equal for AI. It doesn’t have a sense of taste or coherence with a larger ecosystem. I imagine this will be possible in time — I can’t wait for that. But right now, AI is good based on what it’s been trained on. New concepts — like the compose plate or how we present context — are things we’re still refining. We aren’t on our final version of this UI.
The Problem With No Name: Designing Awareness
Kursat: Do you mean that awareness is complicated — the AI knowing what you’re looking at?
Monica: Yes. In a browser, ChatGPT or Gemini web are dedicated apps — they are your endpoint. In Chrome, the focus is on the website, and AI is a helper. That helper must be aware of what you’re looking at and relate your prompts to the page. Expressing that visual relationship quickly and clearly is complicated.
If we spend too much time explaining this, we take away from your goals. We want to ensure transparency so users understand when the assistant is analyzing the page context. Most chat apps have no notion of real-time awareness. We have to communicate how context-awareness works securely across tabs. It shouldn’t get in the way, but it is a new paradigm to introduce clearly. Even if we wanted to design with AI for that, AI wouldn’t be able to invent this kind of thing.
Designing “Voice of the Agent”
Kursat: How do you go about designing agent behavior — the awareness of chat and the window, and the agent clicking around?
Monica: The industry term is “computer use.” An agent creates steps and starts executing them. When we integrated this into Chrome, I was worried about how we signal that a task is about to start. When you ask a question, you could get a text answer or a task that clicks around. I wanted to signal: I’m about to start a task and actuate on your browser on your behalf.
We created “voice of the agent,” blended with Gemini’s task management — visible flags letting people know: I’m a task, and this is what I’m going to do. We iterated on it because people constantly switch between tabs. We needed something glanceable if you ran into a task running. We started using colors; green is good — a basic but effective metaphor.
Because the browser has your credentials, we needed ways for people to feel in control. We won’t log into a website unless you want us to. Transparency and control were key. In the actual browser with your password manager, it’s a bigger deal than on a standalone surface. We took Gemini’s capabilities and added UI to explain: this is a task, you’re in control, and here’s how the browser is clicking itself.
The Glow, Not the Mouse
Monica: The traditional cursor doesn’t actually move. We might simulate movement for visual clarity, but the underlying execution happens at the HTML level. We needed to signal this background automation clearly. A literal cursor simulation proved less optimal in early explorations.
AI has increased the speed of iteration. We focus on rapid prototyping and shipping experimental features to gather feedback early. We initially considered a moving cursor but realized it added complexity without precision. Instead, we introduced a context glow. The window and tab strip feature a subtle glow to communicate background execution smoothly. We aimed to keep it informative and elegant, balancing our product identity with Gemini’s brand language.
Kursat: So, designing an agentic flow has three pillars: the task itself, transparency and control, and the feedback loops. When you’re designing the flow, do you start with the AI’s capabilities and try to invent the paradigms from there?
Monica: Yes. It’s about agentic browsing, but through a user-centric lens. People do many tedious, repetitive things on the internet; if we can save them time, that’s great. We combine our capabilities with where we think the internet is going to evolve. It’s a trio of things: existing capabilities, a real user problem, and industry direction.
The Compose Plate: A New Anatomy
Kursat: What are some emerging interaction design patterns you’re seeing?
Monica: Context and capabilities are at the heart of AI. We’ve been evolving the “compose plate” — the text box where you type. Its anatomy is becoming repetitive across the industry. There’s a plus menu for attachments or capabilities, a model selector, and a submit or live voice button.
It’s an evolution of a search box, but for natural language. It scrolls as you type because people now write full sentences. The compose plate must accommodate that differently than a search query. Two years ago, these weren’t standard, but now they are similar across the industry. It’s your everyday affordance to talk to AI.
Liveliness: The Successor to the Loading Spinner
Monica: Another emerging pattern is showing when AI is thinking or listening — this idea of “liveliness.” In Chrome, we use glows and animations. ChatGPT uses an orb. There’s an awareness and processing that signals the AI is active without needing a literal avatar.
It’s the successor to the loading spinner, but far more expressive. It signals awareness and processing—making the AI feel active and responsive without needing a literal avatar.
Left, Right, or Wherever You Want
Kursat: I know some products put chat on the left and the canvas on the right, and vice versa. How did you approach your panel and canvas design, and what was your rationale?
Monica: If you start in search and keep it as a companion, we put the chat on the left. When Gemini is an on-page assistant meant to spark creativity, we might open it on the right so the page remains the primary focus.
We care about choice. Our AI panels are customizable — you can put them on the right, left, or pop them out. We want to be as flexible as possible for our users.
Agents as a General Contractor, Not a Crowd
Kursat: I’ve been studying with the idea of designing AI agents as teammates — giving them roles, tasks, boundaries, treating this like an organizational design challenge where you intentionally design a team. What’s your point of view on that?
Monica: I use a few tools personally to stay on top of work. I have one agent that scrapes approval items and another that tracks my weekly work. They help me distill topics, and I’ve tweaked their prompts to fit my needs. They mostly handle administrative tasks.
I think I still manage them one by one, to be honest. Some people on the team are really thinking about how to have multiple concurrent agents at the same time. For my own workflow, I get what I need out of them and tweak them individually — I don’t manage them all in one place.
Kursat: When you need multiple agents working concurrently — sequentially, or dependent on each other — are you exploring design work around that? Almost a meta-agent that distributes the work?
Monica: I think if you had a multi-step agent in Chrome, it would probably still behave like a single agent, because it’s really about the jobs to be done. The number of steps could be complex or simple — it could be as simple as adding items to a grocery cart at Walmart, or as complex as researching camping trips and building a list. The complexity of the task changes how many steps need to happen, but it probably wouldn’t change the design much. What happens is Gemini just thinks through the steps, makes a plan, and we execute the plan. So I guess it really is the same, whether the task is simple or complex.
I don’t think I’d want to treat it very differently, because then we’d have to rationalize why it’s different, and explain it. The vast majority of people who are very technical and care a lot about agents understand how the machine is thinking. For people who are novices to agents or AI — the more we can demystify it and make it familiar, the better. If you had a senior agent managing a lot of other agents, I find that would be very complicated for a user — it’s a lot of UI, a lot to explain, and I don’t know if the output would be significantly better or different.
If you hire a general contractor for a big remodel, that’s going to be your main person, and you’d expect the same kind of answers from them as you would from someone doing everything themselves. Now you can have those conversations with the contractor managing everyone else, but the nature of your conversation wouldn’t be very different from one to the other. Because at some point, I think the core of it just goes back to human interaction — how would you interact with another person? Would you want to interact with twenty people doing the job, or just one? You could have twenty, but you could be interacting with just one, because that’s the easiest thing for a person to do.
Kursat: I think we’re now talking about a user experience and an agent experience — from the user level, it doesn’t matter; the user just wants to get their job done. But agent experience could be more technical and complex, something you need to map.
Monica: Yes. It’s important that we understand the mechanics, but expressing that in the UI can raise the barrier to entry. If we make technology details and nuances too transparent, it might be harder for people to actually use it.
Anyone Can Prototype Now
Kursat: I keep hearing about PMs or engineers coming up with a cool prototype and bringing it to design or sometimes attempting to bypass design. Have you had that experience, and how are you navigating the fact that everyone can create a UI or prototype now?
Monica: I think democratic prototyping is fantastic — everyone should be able to express an idea, and new tools make cross-functional collaboration much more expressive. Part of my role is aligning these diverse creative inputs with long-term product strategy to evaluate what we build and why. It’s about synthesizing raw prototypes into a cohesive strategic vision. While anyone can contribute valuable perspectives, our role in design is to curate and ensure product coherence within the broader ecosystem.
Good ideas can come from anywhere. I’m always happy to see people prototype something and share it with me — it happens a lot more now. We get engineers, PMs, even researchers putting together prototypes and ideas, and I think that’s really nice — the flow of creativity and people expressing opinions. Some of them will make it in, some won’t, for different reasons. But the crux now goes back to strategy — how do we curate or shape the product — and I see that as separate from our ability to be generative. Being generative is really good.
As a leader, my job is: how do I receive those ideas, connect with the people who make them, and if I recommend we don’t pursue something, explain why. That’s the important part — explaining why we want to do something, versus why we don’t.
What She Changed Her Mind About
Kursat: What’s one belief about AI and design you’ve changed your mind about after this exposure to integrating Gemini into Chrome?
Monica: It’s super helpful, but my dreams for AI are still bigger than reality yet. It doesn’t always get it right — you can’t really just let it do your job. There’s a lot of fear of AI replacing people, and I don’t know the future, of course, other than what I’ve seen — but not right now. Somebody once made the analogy that AI is like a really smart dog, but it’s not a person. When you find those limits, you’re like, yeah, this thing is not a person. That may change — there are always new developments — but at the moment, it’s still like that.
I thought we’d get there faster, like in the movie Her. But we aren’t there yet. There’s a lot of work to do, and we’ll only get there if we have strong opinions and understand the technology.
Advice for Designers
Kursat: What advice would you give designers navigating this AI wave?
Monica: I think they should use the tools as much as they can. Something I appreciate in designers at Chrome, and in general, is that we want people to have opinions — that’s not new, that’s old. We hire people for their opinions. I assume by the time somebody gets hired, they’re a good designer — the skills are important, maybe especially early on. But once you have the job, the skills continue to matter because they’re how you produce your output, but the quality of your output is also shaped by your opinions. To get good opinions, you have to work hard at them — if your opinions are superficial, easily challenged, or based on intuition you haven’t really thought deeply about, that’s a problem.
When I was at Amazon, they had a leadership principle — “leaders are right a lot” — that at first I didn’t make sense of but later, I understood that: you’re right a lot because you put the work into being right — you think deeply about something, you research it, you consider alternatives, and you form an opinion based on that knowledge. It’s about forming informed opinions, not just relying on intuition.
There’s still a lot that isn’t established yet — the compose plate is narrowing in, but a lot of other things aren’t. When I studied industrial design, we had workshops for ceramics, plastics, wood, metal — because we needed to understand the processes and the materials. I think AI is one of those — you have to spend time in workshop mode, so you really understand the boundaries and the materials.
Kursat: AI is almost a new kind of material we need to learn.
Monica: Yeah — because it’s not deterministic. It has its own weirdness.
Kursat: I think that was a good landing. Thanks so much — I learned a lot, and we teased out some good patterns.
Monica: Thank you for inviting me. This was fun.