David Singleton

David Singleton is a cofounder of Dreamer, a platform for people without technical backgrounds to build and use AI agents. He left Stripe in 2024, where he had been chief technology officer, and joined Hugo Barra in founding /dev/agents, which emerged as Dreamer in March 2026. His work there combined natural-language software creation with developer tools and a central personal agent called Sidekick.[1][2]

On March 23, 2026, Singleton announced that Dreamer’s entire team was joining Meta Superintelligence Labs. The Decoder’s report described the arrangement as licensing Dreamer’s technology to Meta. His March 2026 Latent Space interview, recorded before that announcement, explains the platform’s launch-era decisions.[4][1]

Singleton described his nontechnical sister as an example of the person Dreamer was designed to serve: someone with problems that intelligent software could solve, but no practical way to begin building it. Users could start with community-built applications and ask Sidekick to customize them. The team also hosted the applications, databases and model access, removing the need for users to arrange infrastructure or manage provider credentials.[1]

In Singleton’s account, Sidekick first planned against the agents and tools available, sometimes refusing requests when a necessary tool was missing, then built and tested the application. It also mediated interactions between agents rather than letting them act directly on one another, considering the user’s expectations and tool permissions. He compared this architecture to an operating system: “the sidekicks, like the kernel, the agents and apps are like users.”[1]

Singleton illustrated the approach with a conference planner he built using the prior year’s AI Engineer schedule data, which swyx had made available in machine-readable form. The application let him select sessions and generate a personalized schedule using his interests and Sidekick’s memory of him. He estimated spending 25 minutes on the build across a couple of hours between meetings—a report about that particular application.[1]

For a trip to Big Sky, he built a shared application combining live lift conditions from the community-built Ski Bum tool, accommodation details and dining information supplied by a friend. The group entered expenses and used the application to calculate who owed whom. It did not transfer money: “We didn’t move the money on Dreamer.”[1]

Singleton’s team built Dreamer’s SDK and command-line interface first. Human developers could use the CLI to work locally, and Sidekick used exactly the same interface in coding mode. He argued that good documentation supplied agents with useful context while also improving the experience for human developers. For built-in tools, he described a commitment to stable interfaces while the team updated the models behind them, allowing applications to benefit from improvements without changing their integrations.[1]

The company began writing Python, then built most of its stack in TypeScript, which also became the default for generated applications. Singleton favored TypeScript for its available model training data and compile-time feedback: coding agents could see errors and correct them in a loop. He described type safety extending from the database to the frontend. The SDK still allowed developers to build in other languages.[1]

Other choices reflected experiments and expediency. Singleton said early memory based on a vector database and embedding retrieval worked, but added more complexity than necessary, so the team replaced it. They had also tried knowledge graphs; the system used at interview time was not one. Dreamer’s application versioning was separate from Git because early versions needed to be easy for Sidekick to manipulate. He acknowledged path dependence and said the team might choose Git if starting again; it already used Git for the platform’s own source code.[1]

Singleton advocated small teams because larger groups accumulated communication overhead. At the March 2026 interview, he reported that roughly six people had built the demonstrated core platform and that the company had about 17 employees. His hiring process tested both coding fundamentals and collaboration with coding agents: candidates worked on short projects with an engineer, revealing their product sense, prompting approach and workflow while agents ran. His own workflow included multiple agents, sometimes assigning one to produce code and another to review it.[1]

He considered taste, creativity and individuality continuing human contributions to software development. Dreamer’s team put its own taste into templates, prompts and the harness to reduce generic-looking output. In his assessment, models did not supply those qualities out of the box, even as their capabilities in code, images and music improved.[1]

Before founding Dreamer, Singleton worked on Android at Google and led engineering at Stripe. He explicitly connected Dreamer’s accessibility and developer ecosystem ambitions to his earlier Android experience, while describing production agent work at Stripe as strengthening his conviction about the technology.[1][6]

Singleton ran Android Wear as a general manager. On May 18, 2016, he authored the Android Wear 2.0 Developer Preview announcement as Google’s VP of Engineering. The team’s preview enabled watch applications to access the internet without depending on a nearby paired phone. It also introduced a complications framework through which applications could supply small pieces of information to compatible watch faces, separating the provision of data from its display.[1][5]

At Stripe Sessions 2024, Sharmeen Browarek Chapp introduced Singleton for a usage-based billing demonstration. Using the hypothetical company Hypernian’s AI translation service, he showed how to report word counts through Stripe’s Meters API. A forthcoming Stripe-aware GitHub Copilot extension, trained on Stripe documentation and examples, generated integration code; he then ran a load test and displayed live usage. He also demonstrated trigger-based Automations for enrolling new customers and crediting their balances. These were his public demonstrations of the team’s tools.[2]