Summary

Dreamer’s core bet is that personal agents require an operating system, not a collection of isolated AI apps. David Singleton casts Sidekick as the kernel and agents as users: without that core, vibe-coded apps “grab all your data willy-nilly” and cannot safely cooperate at scale.
The consumer wedge is discovery and immediate utility, with natural-language building available once users want customization. Dreamer targets people like Singleton’s nontechnical sister: install community-built agents, let Sidekick personalize them, and avoid databases, API keys, model tokens, deployment, and authentication. Swyx’s telling observation was that Dreamer puts “discover” before “build.”
The strongest platform flywheel combines third-party tools, remixable agents, and creator economics. Tool builders receive usage-proportional payments; premium integrations such as Parallel Web Systems can charge per use; agent monetization is being considered; and a Builders in Residence program will pay selected creators. Dreamer also offered $10,000 for the best tool added by mid-April.
The demos make a concrete case for bespoke, episodic software whose creation cost is low. Singleton built a personalized AI Engineer conference app in 25 minutes of wall-clock time, using an LLM for every schedule slot and producing recommendations in 30–40 seconds; another app served only a Big Sky ski weekend. His framing: software can finally fit “this one thing” without requiring an expensive agency-built app.
Dreamer is accumulating an integration and orchestration layer above rapidly changing models. Its first-party tools expose stable interfaces for Gmail, search, translation, speech, images, recipes, and direct live sports feeds, while continual evaluations choose among models such as Opus 46, Open AI’s 54 model, and Haiku 4.5. Users get the “right stuff” without tracking the model market themselves.
Inter-agent trust is both the product architecture and the central trust proposition. Agents never call one another directly; Sidekick mediates requests using the user’s permissions, expectations, and memory. The best example was a Granola meeting commitment becoming a task, then a self-completing to-do agent invoking a recruiting agent to make the promised introduction.
Commerce begins with ordinary platform payments, while agent-to-commerce standards remain deliberately unsettled. Stripe Connect already supports paying builders, and some tools move money to close real-world tasks; Dreamer is staying protocol-agnostic while different protocols are evaluated. Singleton compared the moment to about 1995 or 1996, when HTTP coexisted with Gopher and UUCP, and called today’s Overton window “very wide.”
AI leverage is already supporting a small company and changing hiring, but it has not solved taste. Roughly six people built Dreamer’s core, the company has grown to about 17, and interviews now test multi-agent workflows using Code X or Cloud Code. Yet models still default to recognizable “AI generic slop”; Singleton’s remaining frontier is “taste, creativity, sense of individuality,” not another easily verified coding benchmark.

Deep dive

1. Dreamer treats personal agency as an operating-system problem

Singleton defines Dreamer as a consumer platform where anyone can discover, build, and use agents and agentic apps. Its personal agent, Sidekick, can be given a name and personality, learns the user, helps operate the agents on the platform, and translates the act of building software into conversation.
The target user is Singleton’s smart but nontechnical sister: she has problems that intelligent software could solve but no way to get started with tools like Cloud Code or conventional development. Dreamer lets her install something useful from the gallery immediately, then ask Sidekick to reshape it in natural language.
Singleton traces the strategy to building Google’s early mobile apps and core Android components with co-founders Hugo Barra and Nicolas Jitkoff. Mobile required both foundational infrastructure and an ecosystem whose third-party developers could deliver “way more value” than the platform owner could create alone.
Experience putting some of the very first AI agent systems into production at Stripe gave Singleton the conviction that agents would change how people use computers and phones. His conclusion was not merely to ship an agent, but to solve the approachability, coordination, privacy, and distribution problems preventing adoption “for real people at scale.”

2. The product embeds agents into habits users already have

Onboarding begins as a conversation in which Sidekick learns what matters to the user and helps them discover and build initial agents. The resulting dashboard is personal: widgets expose app interfaces, while a feed reports work performed by background agents when the user was elsewhere.
Calendar Hero researches unfamiliar meeting participants, reconstructs prior interactions, and gathers useful public context. Every day it also generates a briefing inside Apple Podcasts; Singleton scans a QR code once, then listens during his morning drive—evidence for his view that agents should appear “in the other apps that you already use.”
The gallery already contained hundreds of community agents, with personal productivity and information management among the top use cases. Singleton’s own AI-news agent continuously filters the public fire hose for items he would care about, while any installed experience can be remixed rather than rebuilt from scratch.

3. Tools turn Dreamer into an open, monetizable supply layer

Singleton’s base-layer principle is that an AI system is only as good as “the quality of data that it can pull in and the quality of action it can take.” Dreamer therefore seeded real Google Search and Gmail integrations plus direct, live data feeds for Formula 1, MLB, and NFL—not sports data scraped from web pages.
The tool layer is open to outside engineers, and Dreamer disclosed that publishers will be paid in proportion to usage. Community examples include Ski Bum’s live resort conditions; Parallel Web Systems offers premium agentic web research on a pay-per-use basis, with free trials available to agent builders.
The new Builders in Residence program fast-tracks creative applicants from the waitlist and pays a small number of the most creative builders to work directly with Dreamer’s team. Singleton likens the system to Lego: when a resident needs a missing “round steering wheel,” the company can build that primitive and make more creations possible.
Dreamer also offered a $10,000 prize for the best tool added by mid-April. The ambition follows Singleton’s favored platform test, which he thinks is a quote from Bill Gates: participants must create and capture more value than the platform itself; tool payments are live, while monetization for shared agents is being considered.

4. Episodic software becomes viable when creation costs collapse

Singleton built an AI Engineer conference app from the event’s llms.txt file and JSON session feed. It supports speaker search, session selection, and a personalized schedule—capabilities he contrasted with expensive agency-built conference apps that are rarely good and become irrelevant as soon as the event ends.
“Guide me” extracts conference themes, combines selected interests with Dreamer’s memory of the user, and runs an LLM prompt across every time slot. The personalized, conflict-free schedule takes roughly 30–40 seconds to assemble and remains available through the same app on mobile.
The first working version emerged after one Sidekick turn; the finished app took 25 minutes of wall-clock effort spread across a couple of hours. Singleton published it to the gallery, where others can use his defaults, fork it, or tell Sidekick exactly how their preferred conference experience differs.
Sidekick first plans around available agents and tools, refuses requests requiring a missing capability, builds, then tests against the stated goal. Initial builds often require 10–15 minutes, but Singleton argues the self-correcting loop means “the first thing that it creates will work most of the time”; subsequent interface edits are faster.

5. Sidekick is both collaborator and permission-enforcing kernel

A community builder connected Granola to a self-completing to-do list, causing meeting commitments to appear automatically. One commitment involved a recruiting introduction; the to-do agent invoked the builder’s separate recruiting agent and completed it—the cross-agent composition Singleton calls “the dream.”
That composition is mediated, not peer-to-peer. An agent asks Sidekick to use another agent, and Sidekick checks the user’s expectations, granted tools, and interests before permitting the action; Singleton describes it as both companion and “traffic cop.”
His operating-system analogy is explicit: “The Sidekick is like the kernel, and the agents and apps are like users.” Picking off only the app-building layer would produce software that takes data “willy-nilly,” cannot cooperate, and ultimately cannot earn the trust required for broad delegation.

6. Dreamer abstracts models while remaining open on commerce

Dreamer continually evaluates state-of-the-art models so consumers need not choose between “Opus 46” and “the 54 model from Open AI.” First-party tools provide stable interfaces for image understanding and generation, RSS, text-to-speech, recipes, and long-form translation even as the best underlying provider changes.
Swyx frames this as an “agent lab”: not a model trainer, but a routing, evaluation, and subject-domain layer. His caveat is that fixed tool structure could work against increasingly merged multimodal models; Singleton’s response is that adaptive sub-agent harnesses should let Dreamer’s primitives evolve with the landscape.
A Big Sky trip app pulled live lift conditions through Ski Bum, combined them with lodging and dining information, tracked group expenses, and calculated who owed whom. Friends used a shared mobile link without already having Dreamer accounts; money still moved outside Dreamer, leaving an opening for a settlement tool.
Stripe Connect is sufficient for paying creators today, while real-world task completion will increasingly exchange value. Singleton remains agnostic among emerging commerce protocols, comparing it to about 1995 or 1996, when HTTP, Gopher, and UUCP were simultaneously viable before one standard achieved dominance.

7. The consumer abstraction sits atop a general engineering stack

Builders can expose logs, generated files, prompts, static assets, and code, or export a project through Dreamer’s SDK and CLI for work in Cursor or Cloud Code. Sidekick itself uses that same CLI, making complete documentation useful to both humans and coding agents.
Dreamer runs general code in a VM rather than constraining builders to a narrow stack. TypeScript is the default because models have abundant training data for it and strong typing lets coding agents discover mistakes at compile time; Singleton still prefers Python when writing code manually.
Dreamer’s own stack carries type safety from database to front end. Its built-in version system is not Git—Singleton says it can work more efficiently than Git and explains that the early versions were an expedient optimized for Sidekick manipulation—but builders can inspect every historical version. He also says that, if they started over, they might make it Git.
Activity views expose webhooks and email triggers with filters, including LLM filters, as well as individual prompts and model calls; the schedule demo used Haiku 4.5. Each agent receives a multi-user SQLite database, platform authentication, row-level ownership defaults, and exported functions Sidekick can invoke, while deliberately public data must be specified.

8. Memory and agent leverage support a deliberately small company

Sidekick’s accumulated memory lets agents begin with relevant context. A weekend planner recommended the Civic Center St. Patrick’s Day parade because it knew Singleton was Irish, turning a stable personal fact into a timely family suggestion.
Dreamer previously stored facts in a vector database and retrieved them through embeddings and RAG; it worked, but proved more complex than necessary. The team also tried knowledge graphs and “probably implemented most of the papers” on agentic memory before settling on a simpler current system, with more than one person still improving it.
Becoming CEO raised the temperature because “the buck truly stops with you” and expanded Singleton’s attention toward go-to-market and direct user contact. Yet the core product shown was built by about six people; Dreamer, founded in late 2024, had grown to roughly 17 while using its own coding, marketing, and operations agents.
Hiring now emphasizes collaboration with coding agents. Candidates pass a conventional coding screen, then develop a fully formed product idea with an engineer using Code X, Cloud Code, or their preferred system; Dreamer watches their prompting, product sense, supervision, and ability to round-robin multiple agents or chain one coder to another reviewer.

9. Verifiable execution is improving faster than creative taste

Singleton’s answers have changed quickly: a year earlier he would have requested better music generation, and six months earlier better images. Current Suno tracks no longer sound obviously generated to him, while the latest Nano Banana produces “spectacular infographics” and painterly images through a Dreamer tool.
The remaining deficit is human taste: understanding motivation, choosing what delights, and producing individuality. Dreamer invested its own taste into templates, prompts, and the harness to avoid “AI generic slop,” because models still have recognizable default aesthetics—Singleton says he can identify which model built a basic to-do app by looking at it.
Swyx’s pushback is that taste is broad, “I’ll know it when I see it,” and therefore difficult for researchers to verify. Singleton agrees code advanced partly because correctness is tractable, but expects people to start debugging the creative process and thinks some iconic companies could become places for that work; Swyx’s closing measure is that machines may have moved from 30% to 95%, yet the final 5% persists.

This is why I don't believe in sales quotas | Figma CRO Terence Tao – How the world’s top mathematician uses AI