Dreamer: The Agent OS That Treats AI Like Apps

Why Dreamer Frames AI Agents as an Operating System

David Singleton spent years as CTO of Stripe, where his team deployed some of the first production AI agent systems. That experience convinced him that the next computing paradigm shift — agents doing work autonomously — needs the same foundational infrastructure that mobile apps needed in 2008. So he co-founded Dreamer with Hugo Bar and Nicholas Czechov, his colleagues from Google’s early Android team, to build exactly that.

The OS metaphor is literal, not marketing. “The sidekick’s like the kernel, the agents and apps are like users. Different rings.” Dreamer’s core innovation is a personal agent called the “sidekick” that mediates everything. When one agent needs to interact with another, it goes through the sidekick — which understands user permissions, tool access, and intent alignment. This prevents the failure mode Singleton sees in vibecoded apps: “You could build little vibecoded apps, but they’re going to grab all your data willy-nilly. They won’t be able to work together.”

Consumer-first, not developer-first. Singleton designed Dreamer for his non-technical sister, not engineers. The sidekick handles onboarding, helps users discover agents from a gallery, and builds custom apps through natural language. A conference schedule planner, a self-completing to-do list, a financial fitness coach — all built by users in minutes, not weeks.

The marketplace economics are real. Tool builders on Dreamer get paid proportionally to usage — a revenue-sharing model that mirrors the Play Store ecosystem Singleton helped build at Google. Premium tools like Parallel Web Systems operate on pay-per-use, while community-built tools (ski conditions, Formula 1 live data) are free. This creates a genuine two-sided marketplace where agents compose tools into applications.

Agents that escape the chat window. One of Dreamer’s most compelling features is output delivery: agents generate daily briefing podcasts that appear in Apple Podcasts, calendar widgets that research meeting attendees, and background feeds that surface agent activity. “Making things show up in the other apps that you already use in your life is incredibly powerful.”

5 Key Insights from David Singleton on Agent Platforms

• The self-completing to-do list works — Singleton describes users connecting Granola (meeting notes) to a to-do list agent that spawns sub-agents to complete tasks, including using a recruiting CRM agent to execute introductions mentioned in meetings
• Hiring has fundamentally changed — Dreamer’s interview process tests how candidates work with coding agents, looking for “a nice round robin” workflow where engineers review one agent’s output while another works
• Engineering managers make great agent operators — “Being an engineering manager, as long as you stay very close to the code, is actually a great skill profile for being able to make agents work for you”
• Build-test-iterate loops create magic — Dreamer’s agent studio has the sidekick plan, build, and then test its own output, producing working apps on the first build “most of the time” in 10-15 minutes
• Taste is the remaining frontier — Singleton argues LLMs still lack creativity and individuality: “I can tell which model built it just by how it looks.” Dreamer invests heavily in templates and prompts to prevent “AI generic slop”

What Agent OS Design Means for Autonomous Work

Dreamer’s architecture validates a thesis that’s been building across the industry: agents need infrastructure, not just prompts. The sidekick-as-kernel pattern — where a trusted agent mediates permissions, coordinates inter-agent communication, and maintains user context — is essentially the same pattern that operating systems solved for applications decades ago. For organizations building agent platforms, the lesson is clear: security, composability, and trust aren’t features to bolt on later. They’re the foundation that makes everything else possible.