Most important take away

Dreamer is positioning itself as the “operating system for AI agents” — a consumer-facing platform where non-technical users can discover, build, customize, and share agentic apps using natural language, while engineers can contribute tools and monetize them. The core insight is that the agent ecosystem needs the same kind of platform infrastructure (auth, databases, hosting, inter-agent communication, privacy) that mobile needed in Android’s early days, and the team that built Android’s app ecosystem is now applying that playbook to AI agents.

Summary

Actionable Insights:

• Try Dreamer’s builder-in-residence program: Visit dreamer.com/latent-space to skip the waitlist and potentially get paid to build agents on the platform. Tool builders also get paid proportionally to usage. There is a $10,000 prize for the best tool submitted by mid-April 2026.
• Build and publish tools on Dreamer: The platform is open for third-party tool builders. If you publish a tool that agents use heavily, you earn money. This is an early ecosystem where first movers can establish themselves — think early App Store or Play Store dynamics.
• TypeScript over Python for AI agent code: David Singleton, former Stripe CTO, argues TypeScript is superior for AI-generated code because strong typing lets coding agents catch errors at compile time, creating a feedback loop that dramatically improves output quality. Dreamer’s entire stack uses TypeScript with type safety from database to frontend.
• Career advice — work effectively with coding agents: When hiring engineers, Dreamer evaluates how well candidates work with coding agents (Claude Code, Codex, etc.), including prompting strategy, what they do while agents work, and ability to run multiple agents simultaneously in a round-robin fashion. Former engineering managers who stayed close to code are particularly well-suited for this new workflow.
• Career advice — small teams with high talent density: The core Dreamer product was built by approximately six people. Singleton strongly advocates for small teams because communication overhead grows nonlinearly with headcount. Every person at the company is an individual contributor, including former managers.
• Write great CLI documentation: Dreamer’s sidekick coding agent uses the same CLI that human developers use. The implication for the industry: making your CLI well-documented and agent-friendly is now a competitive advantage. Singleton specifically called out Stripe needing to do this.
• Memory and personalization as a moat: Dreamer builds a persistent user profile through its sidekick agent. They tried vector databases with RAG embeddings and knowledge graphs but moved to a simpler system. The key takeaway is that the more you use the platform, the more personalized and useful it becomes — this is their retention moat.
• No stocks or investments were explicitly mentioned as recommendations. Dreamer is a startup (founded late 2024, roughly 17 employees). They use Stripe Connect for payments. The broader thesis discussed is that the “agent commerce” space is in a Wild West phase similar to early web protocols competing — no clear winner yet among A2A, MCP, and similar protocols.

Chapter Summaries

Introduction and What Is Dreamer (0:00) David Singleton, former Stripe CTO, introduces Dreamer as a consumer-facing platform where anyone can discover, build, and use AI agents and agentic apps. The company was co-founded with Hugo and Nicholas Chekov, all of whom worked together on early Android at Google.

Live Demo — Conference App (Mid-early) Singleton demos a conference scheduling app he built for AI Engineer World’s Fair in about 25 minutes by talking to the sidekick. The app pulls speaker data, lets users pick sessions, and uses LLM calls to generate a personalized schedule based on user interests and sidekick memory.

Platform Architecture — Tools, Gallery, and Sidekick (Mid) Dreamer has three layers: tools (integrations like Gmail, Google Search, Formula One data, ski conditions), a gallery of community-built agents, and the sidekick which orchestrates everything. Tool builders get paid, premium tools exist on a per-use basis, and the platform handles auth, databases (SQLite per agent), and hosting automatically.

Agent Studio and Developer Experience (Mid-late) The agent studio lets users build via natural language conversation with the sidekick. Under the hood, there is a full IDE with build logs, file browser, prompt editor, and version history. An SDK with CLI allows engineers to pull code locally and edit in Cursor or Claude Code. Everything is TypeScript-based.

Security, Privacy, and the OS Analogy (Mid-late) The sidekick acts as a kernel — all inter-agent communication goes through it. It enforces permissions and ensures agents only access data they are authorized to use. Multi-user apps handle data isolation automatically unless the builder explicitly makes data shared.

Memory and Personalization (Late) The sidekick builds persistent memory about users over time. They tried vector databases with RAG and knowledge graphs but settled on a simpler system. Multiple team members work specifically on memory. Example: a weekend activity planner that recommended an Irish parade because it knew the user is Irish.

Company Building and Hiring (Late) Dreamer has about 17 people with very high talent density. The interview process includes a traditional coding screen plus collaborative sessions building real products with coding agents. They evaluate how candidates prompt agents, manage multiple concurrent agents, and demonstrate product sense.

The Future and What LLMs Still Lack (End) Singleton says the biggest gap in LLMs is taste and creativity — AI-generated apps look generic without significant human curation. Dreamer invests heavily in templates and prompts to avoid “AI slop.” He sees quantifying and teaching taste as the next frontier for AI research.