An AI agent that flows with you
How we created an AI agent that doesn't get in your way and helps you work better.
Role
Product designer
Proyect
Brave AI
Year
2023-Ongoing
In the rapidly evolving landscape of AI chatbots and assistants, we at Brave faced a big challenge:
How do we create an AI companion that enhances productivity while staying true to our core values of privacy and security?
This case study explores how the team tackled this challenge by designing Leo, an AI assistant that works alongside users rather than demanding their constant attention.
In the rapidly evolving landscape of AI chatbots and assistants, we at Brave faced a big challenge:
How do we create an AI companion that enhances productivity while staying true to our core values of privacy and security?
This case study explores how the team tackled this challenge by designing Leo, an AI assistant that works alongside users rather than demanding their constant attention.
The challenge
The fundamental problem of AI assistants is that they promise to save us time, yet often require significant oversight. When we began designing Leo, we observed that existing AI assistants created a "productivity tax" – users had to stop their work to supervise the AI, defeating the purpose of delegation. Additionally, as a privacy-first browser, we faced technical constraints that many of our competitors didn't have to consider.
Working with a small development team meant we needed to be strategic about our feature prioritization. We couldn't compete with tech giants on raw computational power, so we had to be smarter about how we utilized our AI resources.
Design Process and Solutions
Our breakthrough came when we reframed the problem: Instead of trying to create a perfect AI assistant, we designed for imperfection. This led to several innovative solutions:
Parallel Productivity
We developed a unique "multiplayer" browsing experience where Leo gets its own cursor, allowing users to continue their work while monitoring Leo's progress
in their peripheral vision. It's like pair designing/programming, but with an AI partner that doesn't mind if you look away
The challenge
The fundamental problem of AI assistants is that they promise to save us time, yet often require significant oversight. When we began designing Leo, we observed that existing AI assistants created a "productivity tax" – users had to stop their work to supervise the AI, defeating the purpose of delegation. Additionally, as a privacy-first browser, we faced technical constraints that many of our competitors didn't have to consider.
Working with a small development team meant we needed to be strategic about our feature prioritization. We couldn't compete with tech giants on raw computational power, so we had to be smarter about how we utilized our AI resources.
Design Process and Solutions
Our breakthrough came when we reframed the problem: Instead of trying to create a perfect AI assistant, we designed for imperfection. This led to several innovative solutions:
Parallel Productivity
We developed a unique "multiplayer" browsing experience where Leo gets its own cursor, allowing users to continue their work while monitoring Leo's progress
in their peripheral vision. It's like pair designing/programming, but with an AI partner that doesn't mind if you look away
Usage of APIs
Most of the websites we interact with daily offer a set of APIs that allow other apps to interact with them on the background. We wanted to make this the primary way that Leo interacted with a website. This would allow us to reduce the chances of errors, while also saving a lot of time, since tasks can now be made almost instant and in the background.
In this example, you want to buy tickets for a show, but you're not sure if you have anything planned on that day already. Instead of interrupting your flow to check, you can just ask Leo to check your calendar for you and tell you. It all happens in the background (considering you've already granted access to your Google account) and instantaneously.
Smart progress visibility: embracing human time
Now, APIs don't solve all use cases, so we also needed to fall back into the more "classic" AI agent routine, where Leo takes over the browser and clicks around to figure stuff out.
The sidebar became our solution to the "black box" problem of AI assistants, but it also helped us tackle a deeper challenge: setting the right expectations about AI working speed. Rather than pursuing instantaneous responses -which often lead to errors- or leaving users in the dark with slow processes, we designed a system that embraces "human time" - the natural pace of real-world browser usage.
Think about how a human assistant would book a restaurant table: they'd need to look up the website, navigate to reservations, check availability, and fill out forms. This is because the web is built for human usage. We need to see actions as multiple steps, dropdowns, input fields, options. Each step takes a few seconds, and that's perfectly natural for us.
We designed Leo to work at this same comfortable pace, making its actions both more reliable and more comprehensible.
The persistent progress indicator keeps users informed without demanding their attention, similar to how you might glance at a colleague's screen to see how they're doing on whatever they'd be working on.
This human-paced approach serves multiple purposes:
It allows Leo to work more accurately by not rushing through complex tasks
It gives users natural moments to intervene or redirect if needed, without the sense of FOMO of not being able to digest the AI actions in time
It sets realistic expectations about AI capabilities
It makes the assistance feel more natural and trustworthy
The result is an experience that feels less like issuing commands to a machine and more like working alongside a capable colleague - someone who works thoroughly rather than rushing, and keeps you informed along the way. Just as you wouldn't expect a human assistant to instantly complete a complex task, Leo's measured pace helps build a more authentic and reliable collaboration.
Trust through control
We implemented a confirmation-based system for complex tasks, but with a twist. Instead of requiring constant supervision, Leo breaks down complex tasks into digestible checkpoints. Users can review these at natural breaking points in their own work, maintaining flow while ensuring accuracy.
Usage of APIs
Most of the websites we interact with daily offer a set of APIs that allow other apps to interact with them on the background. We wanted to make this the primary way that Leo interacted with a website. This would allow us to reduce the chances of errors, while also saving a lot of time, since tasks can now be made almost instant and in the background.
In this example, you want to buy tickets for a show, but you're not sure if you have anything planned on that day already. Instead of interrupting your flow to check, you can just ask Leo to check your calendar for you and tell you. It all happens in the background (considering you've already granted access to your Google account) and instantaneously.
Smart progress visibility: embracing human time
Now, APIs don't solve all use cases, so we also needed to fall back into the more "classic" AI agent routine, where Leo takes over the browser and clicks around to figure stuff out.
The sidebar became our solution to the "black box" problem of AI assistants, but it also helped us tackle a deeper challenge: setting the right expectations about AI working speed. Rather than pursuing instantaneous responses -which often lead to errors- or leaving users in the dark with slow processes, we designed a system that embraces "human time" - the natural pace of real-world browser usage.
Think about how a human assistant would book a restaurant table: they'd need to look up the website, navigate to reservations, check availability, and fill out forms. This is because the web is built for human usage. We need to see actions as multiple steps, dropdowns, input fields, options. Each step takes a few seconds, and that's perfectly natural for us.
We designed Leo to work at this same comfortable pace, making its actions both more reliable and more comprehensible.
The persistent progress indicator keeps users informed without demanding their attention, similar to how you might glance at a colleague's screen to see how they're doing on whatever they'd be working on.
This human-paced approach serves multiple purposes:
It allows Leo to work more accurately by not rushing through complex tasks
It gives users natural moments to intervene or redirect if needed, without the sense of FOMO of not being able to digest the AI actions in time
It sets realistic expectations about AI capabilities
It makes the assistance feel more natural and trustworthy
The result is an experience that feels less like issuing commands to a machine and more like working alongside a capable colleague - someone who works thoroughly rather than rushing, and keeps you informed along the way. Just as you wouldn't expect a human assistant to instantly complete a complex task, Leo's measured pace helps build a more authentic and reliable collaboration.
Trust through control
We implemented a confirmation-based system for complex tasks, but with a twist. Instead of requiring constant supervision, Leo breaks down complex tasks into digestible checkpoints. Users can review these at natural breaking points in their own work, maintaining flow while ensuring accuracy.
Looking forward
Looking into the near future, tasks that initially could be done by the agent browsing in real-time can be made instant, by taking advantage of APIs, like in this example above, where I can just ask Leo to give me a list of all the things that happened since I stopped working the previous day, ready by the time I start back up today.
As AI technology continues to evolve, we're excited to further refine Leo's capabilities while staying true to our core design principles. The future of AI assistance isn't about creating a perfect autonomous system, but about designing tools that enhance human capabilities while respecting autonomy.
More projects
CASE STUDY
Brave's Design System
Creating Nala: One design system to rule all our platforms and products.
CASE STUDY
Reimagining the browser for productivity
Creating a set of features to increase productivity in the browser.
Custom browser icons
Adding fun ways to customize your home screen experience.
Leo AI webapp
Creating a more comfortable UI for deep work with AI.
Contextual conversations with Leo AI
How we created an UX for easily adding the context of a webpage to an AI conversation.