Dreamer: the Agent OS for Everyone — David Singleton
TL;DR
Dreamer is pitched as an "Agent OS," not just an app builder — David Singleton says the hard part isn’t generating little vibecoded apps, it’s building the operating-system-like core where the Sidekick acts like a kernel, manages permissions, and safely coordinates agents, tools, and user data.
The product is explicitly consumer-first, but the stack underneath is deeply technical — Dreamer starts with discovery for non-technical users like Singleton’s sister, yet under the hood it includes an agent studio, SDK, CLI, build logs, prompts, exported functions, triggers, and a multi-user SQLite model that engineers can inspect or pull local.
Dreamer’s killer demo is bespoke software made in minutes for real-life moments — Singleton built an AI Engineer conference planner in about 25 minutes using Latent Space’s LLM.txt and JSON feeds, and also made a Big Sky ski-trip app with live lift data, shared expenses, and group settlement for one weekend with friends.
Tool builders are central to the platform economics — Dreamer already has hundreds of community-built agents, pays tool builders based on usage, offers premium-tool trials, launched a builders-in-residence program, and announced a $10,000 prize for the best new tool added by mid-April.
The company’s thesis is that routing, evals, and integrations should be invisible to end users — Dreamer chooses among models like Claude Haiku or OpenAI variants, keeps first-party tools like translation and image generation stable while swapping backends as models improve, and treats this “agent lab” layer as necessary if normal people are going to use AI at scale.
Singleton thinks small, agent-augmented teams are the new company shape — Dreamer shipped the core product with roughly six people, is now around 17, runs many internal workflows on its own agents, and hires engineers partly by watching how they collaborate with multiple coding agents in parallel, not just how they type code by hand.
The Breakdown
From Android and Stripe to an "app store for agents"
Singleton introduces Dreamer as a place where “literally everyone” can discover, build, and use AI agents, with a special emphasis on non-technical consumers. He frames the company as a direct sequel to lessons from Android and Google Play with co-founders Hugo Barra and Nicholas Jitkoff, then sharpened by seeing some of the first production agent systems at Stripe and realizing this would change how people use computers.
The first demo: your Sidekick, your dashboard, your daily podcast
Once you’re in Dreamer, everything starts with a personal agent called Sidekick, which gets to know you and helps you use or build agents across the platform. Singleton shows his own dashboard: a feed of background agent activity, widgets like Calendar Hero, and a daily briefing podcast generated by an agent and pushed straight into Apple Podcasts via QR code — something he now listens to in the car each morning.
Gallery, tools, and the paid ecosystem underneath
He walks through the three-layer structure: installable agents from the gallery, a tools layer with integrations like real Gmail and Google Search, and Sidekick stitching it together in the agent studio. The notable reveal is economic: third-party tool builders can publish tools, get paid proportional to usage, and even offer premium tools like Parallel Web Systems, while Dreamer seeds the ecosystem with high-quality data feeds like live MLB, NFL, and Formula 1 rather than scraped web data.
The conference app that took 25 minutes
Singleton’s favorite proof point is a custom app for the AI Engineer conference, built from Latent Space’s structured site data. It lets you search speakers, add sessions, then hit “guide me” so the system reads themes like agents, code generation, and reasoning/RL and creates a personalized schedule in 30 to 40 seconds — the whole app took him about 25 minutes of wall-clock time between meetings.
How building works: plan, code, test, host
Dreamer’s coding flow is intentionally slower up front — often 10 to 15 minutes for a first build — because Sidekick first plans the architecture, then writes it, then tests it in a loop until it actually works. Singleton emphasizes that this is where “magic happens”: coding agents improve dramatically when they can inspect their own output, and Dreamer exposes prompts, code, logs, files, and even a CLI/SDK for technical users who want to take the generated app into Cursor or Claude Code.
Agents that complete each other’s work
One of the most vivid moments is the self-completing to-do list example, which Singleton notes is also Sam Altman’s “number one ask” for AI apps. A builder connected Granola so spoken commitments in meetings become to-dos automatically, and those to-dos can trigger other Dreamer agents — like a recruiting CRM agent that actually makes an intro — which is the closest the conversation gets to the company’s real ambition: software that quietly advances your life across apps.
Why Dreamer had to become an operating system
This is the core thesis of the interview. Singleton says privacy and trust make the architecture non-optional: agents cannot just talk directly to each other and grab data “willy-nilly,” so Sidekick becomes the kernel-like traffic cop that mediates permissions, actions, and user intent across the system, like apps and users operating in different rings of an OS.
Company building, memory, and the taste frontier
In the final stretch, Singleton talks about Dreamer as a tiny, high-talent-density company: the core product was built by about six people, and the team is now roughly 17. He says memory is a major investment area — they tried embeddings, vector search, and knowledge graphs before settling on a more practical system — and ends on a more philosophical note: models are getting astonishingly good at code, music, and images, but they still don’t have taste, which he sees as the next hard frontier in making AI-generated software feel less like generic slop and more like something a person would actually love.