Cuervo café
PEOPLE ARE ORDERING
Raw zucchini, zucchini and pistachio pesto
PEOPLE ARE SAYING
Highly praised for its wide wine selection—by the glass and bottle—focused on boutique producers, plus creative, seasonal small plates with great vegetarian options.
News/Feed
The US public's frenzy over Jackie O's possessions
bbcKonate says he is close to signing new Liverpool deal
bbcDad brains: How fatherhood rewires the male mind
bbcThe US public's frenzy over Jackie O's possessions
bbc
10 PMYou have a very long day ahead
01.
Some shifts in computing are visible only in hindsight. Others unfold so quickly that by the time we begin describing them, the industry has already moved on.
Agentic software may be one of those shifts. Over the past year, the conversation around AI has moved from models to systems. From copilots that assist users to agents that act on their behalf. What felt experimental not long ago is already becoming a design constraint for new products. Most teams are still catching up. Over the past months, we got to explore it from the inside.
The question is no longer whether software will act on your behalf. It's what that means for how we design it.
The question is no longer whether software will act on your behalf. It’s what that means for how we design it.
The question is no longer whether software will act on your behalf. It’s what that means for how we design it.
02.
Dreamer is the AI-native operating system being built by the team at /dev/agents. Instead of treating agents as features inside applications, Dreamer treats them as first-class entities of the system itself. Persistent, contextual, and capable of operating across multiple environments.
The founding team includes David Singleton, former CTO at Stripe and VP of Engineering at Google, and Hugo Barra, who helped scale Android during its early years before leading at Xiaomi and Meta. Around them: a group of engineers and designers who shaped foundational systems at Google, from Material Design and Chrome OS to the infrastructure behind Search and Translate. In 2024, /dev/agents raised $56 million in seed funding, led by Index Ventures and CapitalG, with backing from Conviction and angels including Andrej Karpathy, Alexandr Wang, and Andy Rubin.
The platform they built gives every application a base that users can take, modify, and reshape into something that fits their own life. Start with a news briefing and add a layer that filters by your calendar. Take a restaurant recommendation app and tune it around your neighborhood, your diet, your schedule. It is, by design, infinitely reconfigurable.
What makes building and designing for Dreamer different isn't the agents. It's Remix. In Dreamer, apps aren't finished products. They're starting points.
Just a sec.
We’re gathering the latest stories for you.
This architecture is supported by Sidekick, Dreamer’s integrated AI assistant, which helps users understand how their apps behave and connects them to the outside world through tools and MCPs. Search the web, read emails, call external APIs. The bridge between the agent’s intelligence and any service the user needs.
The apps we shipped would define what was possible before anyone else had a chance to.
03.
Hugo framed the work from day one.
When he laid out what he needed from us, he reached for a specific analogy.
The Super Nintendo. The first titles released for the platform were more than games. They defined what quality looked like for everything that followed. Super Nintendo had Mario. Game Boy had Tetris. Those weren't just launch titles. They were the standard-setters. Every developer who came after had to answer the same question: how do we measure up to that?
His point for Dreamer was the same. The first apps set the standard. Make them exceptional, and that becomes the new floor for the entire ecosystem.
That was the only brief we needed.
See the how it
looks on you!
Sidekick
Okey, let’s get building. What should we do?
You
I want to create an app to try on outfits digitally, but starting from your own photo
Sidekick
Let’s go deeper
You
First, it should transform any casual photo into a clean studio-style image, good lighting, white background, neutral pose… Users should be able to adjust their appearance before trying look.
Sidekick
Got it.
Sidekick
Removed from saved looks
This look is no longer in your saved collection. Changed your mind? Use the Save button to add it back before closing.
Link copied to clipboard!
Your looks
Everything you’ve saved, in one place.
Hey there!
What do you want to try on today?
Your saved look
Onboarding
User data analysis
Start with a photo
Fix photo
Choose an outfit
Create
04.
Daily news briefings, personalized podcasts, cultural discovery tools, weekend planners, restaurant recommendations, climate visualizations, virtual try-on experiences.
We went in thinking we were shipping apps. What we were actually doing was running experiments.
Each one was a question: what happens when software remembers your preferences over time? When it reads your calendar, your location, your past decisions, and decides what to surface next? When it acts before you ask?
Traditional sprint cycles felt slow almost immediately. Planning weeks ahead made no sense when the platform itself was changing in real time. So we adapted as we went. Ideas moved from concept to working prototype in two or three days. Decisions that would normally live in a kickoff doc got made in the middle of a build. The rhythm compressed in ways we hadn't anticipated, and we had to find a way to move fast without losing the thread.
We used the first app, a weather visualization, as a calibration tool: to understand what working in this environment actually demanded, where the platform's limits were, where our own limits were, and what process could hold up across the six apps that followed.
05.
Early on, the classic split between design and development broke down fast. Working in two separate tracks created a gap that, at this speed, we couldn't bridge cleanly. By the time something crossed from concept to implementation, context had already shifted. The platform had moved.
So we dismantled the model entirely.
Designers started coding their own work. Developers started making design decisions while building. Claude Code, Figma Make, and Sidekick became the infrastructure of the workflow, not tools layered on top of it. Claude Code for implementation and iteration. Figma Make for collapsing the distance between wireframe and build. Sidekick as a live QA partner inside the platform, helping us understand how each app actually behaved in context.
Working this way required a specific kind of discipline. Vibecoding with speed is straightforward. Vibecoding with pixel-perfect thinking is something else.
The craft had to stay intact even as the cycles compressed. Sometimes the AI took the work somewhere better than what we'd originally designed, and knowing when to follow that direction, rather than override it, became a skill in itself.
06.
The hard part of each app was never the interface. It was the logic driving it. We knew we were designing agents, not applications. What we didn't fully anticipate was how deep that distinction ran.
The first time we designed a news briefing aware of a user's calendar, the instinct was to treat it like a feature: add the data, surface it in the UI. It worked. But it felt flat. What we hadn't answered yet was how the system should decide what mattered, when, and why.
That pushed us to think about concept differently. In a traditional product, identity lives in the visual layer. Here it had to live in the behavior. We started calling it conceptual identity: not just how an app looked, but how it made decisions. The logic that determined what to surface, what to ignore, and why, across every possible state and context.
Validating also shifted. Wireframes couldn't show us what we needed to see, so we stopped using them as the primary feedback loop. That compression sharpened our taste fast. At that pace, it was the differentiator. Each app ended up with its own identity.
Built for a system that behaves rather than just displays, and that anyone could take as a starting point and make their own.
Add aNew City
07.
Design and development have to move as one.The handoff model is too slow and too lossy for this kind of work. The people who move fastest are the ones who can hold both in their heads at the same time, make the call with taste, and keep going.
The team is the process.When there's no playbook, clarity of direction matters more than structure. The PM's job stops being about managing tickets and becomes about keeping the team unblocked and pointed at the right thing. That's the compass.
AI belongs inside the workflow, not alongside it.On this project, Claude Code, Figma Make, and Sidekick were the connective tissue. The infrastructure that kept us fast without sacrificing the craft.
Speed and craft aren't a tradeoff.Compressing cycles only works if the obsession with detail stays intact. Move fast with pixel-perfect thinking, and velocity becomes your edge.
This project changed how we work, and we're not going back.
A platform that didn't exist when we started and looked different every week while we were building on it. What we shipped was a way of working. One we're taking into everything we do next.
When every user can take an app and make it entirely their own, the question for us stopped being what should this look like. It became something way more interesting: what should this be capable of becoming?
