Dreamer: Building a Personal Agent Operating System

Start with someone who cannot build the app

David Singleton starts with his sister: she has problems that useful software could solve, but no technical background and no obvious way to begin. [Dreamer](https://dreamer.com/latentspace) gives her a starting point—a community gallery of apps she can use and adapt. Its central personal agent, Sidekick, helps operate those apps and build new ones, with a name and personality the user chooses.

Singleton connects the approach to his early Android work with co-founder Hugo Barra: solve the platform problems that make a breakthrough usable, then enable third-party developers to deliver more than the platform team could build alone. Deploying early agent systems at Stripe gave him a similar conviction about AI—and an appreciation of the work needed to make it approachable. At the time of this Latent Space conversation, Dreamer had recently entered beta, with access through a waitlist.

A meeting briefing in Apple Podcasts

Onboarding begins with a conversation about the user’s interests and needs. Sidekick helps discover or build initial agents, then presents a personalized dashboard with widgets and a feed of background activity. Calendar Hero, Singleton’s first example, researches unfamiliar meeting participants and combines public information with prior interaction history.

Calendar Hero also generates a daily briefing podcast. Singleton shows it in Apple Podcasts, connected by scanning a QR code, and says he listens during his commute. Dreamer supports voice conversations with Sidekick on mobile, but the briefing illustrates another useful interface: deliver completed work into an app someone already uses.

The gallery reportedly contained hundreds of agents, especially for productivity and information management: documents, email, and filtering public information. Singleton uses one to monitor AI news and select relevant items. Users can modify installed experiences through natural-language requests to Sidekick, making discovery a route into customization.

Tools supply the data and actions beneath these apps. Dreamer seeded Google Search and Gmail integrations before beta; sports requests from alpha users prompted Formula One, MLB, and NFL tools backed by direct live feeds. External developers can contribute capabilities such as Ski Bum, which supplies resort snow conditions. Published tool builders are paid in proportion to agent usage, though rates and payout terms were not specified.

Some partner tools charge per use, including Parallel Web Systems for agentic web research. Premium tools offer free trials so builders can test whether a capability serves their use case before subscribing.

A conference planner built around its attendee

Conference apps can be expensive to build yet useful for only a brief event. The next demonstration addresses that mismatch with a personal AI Engineer schedule planner, built through conversation using the previous year’s conference data. An llms.txt file provided machine-readable information, alongside a JSON session feed. The resulting app supports speaker search, session participants, and manually adding sessions to a schedule.

Its Guide Me flow fills out the schedule: 1. Extract themes from the conference program. 2. Ask the user to select interests. 3. Run a separate large language model (LLM) prompt for each time slot, choosing a session using those interests and Dreamer’s memory of the user. 4. Display the selections in a calendar. Generation was estimated to take 30–40 seconds during the demonstration. The resulting schedule was described as conflict-free—a single demonstrated outcome, rather than a measured reliability rate.

The demonstrator estimated spending 25 minutes building the app, spread across a couple of hours between meetings. The planner was also shown working on a phone and published to the gallery, where others could fork it and request changes. This complements the conference organizer’s rationale for publishing machine-readable data: a minimal official app can support many independently tailored experiences.

Plan, build, test, repair

The planner’s first prompt specified its purpose and data source; a simple working version reportedly emerged after one Sidekick turn. Agent Studio supports this conversation on desktop and phone. It translates output from the agent harness—the software running the coding workflow—into readable English in Sidekick’s configured personality, while keeping technical logs available behind an optional view.

Sidekick first examines existing agents and tools, then plans how to obtain the required data and capabilities. It can refuse a request and identify a missing tool. Once it builds the app, it tests whether the result fulfills the goal, inspecting output and repairing bugs in a loop. Initial builds reportedly often take 10–15 minutes, with smaller changes completing faster; the claim that first builds usually work comes without a measured success rate.

An integrated development environment exposes generated code and prompts, including the planner’s session-selection prompt, for direct editing. Dreamer handles hosting, database provisioning, and model access. Sharing has two distinct meanings: - Publish a reusable app to the gallery for others to adapt. - Share a running instance with friends or coworkers.

Sidekick coordinates work across apps

Dreamer uses agents internally, including one that researches waitlist signups to prioritize prospective nontechnical builders likely to provide feedback. Any app can launch a Sidekick task: a sub-agent running in its own virtual machine (VM), with Sidekick’s context and access to platform tools and public web data. Researching people through public information is a common use case; a ZoomInfo integration was not available at the time.

A community-built self-completing to-do list launches a task for each item. It may finish the work, invoke another agent, or research enough to help the user begin. One builder connected Granola so meeting commitments became tasks automatically. A promised recruiting introduction reportedly moved from meeting notes to the list, which used a recruiting agent’s candidate and contact-management capabilities to complete it.

This coordination depends on generated apps exposing their internals to Dreamer. Agent-to-agent work passes through Sidekick, which is intended to apply the user’s expectations and tool permissions. That is the substance of the operating-system analogy: Sidekick plays the kernel’s coordinating role, while apps rely on a shared core for interoperability and controlled data access.

Paying builders and completing real-world tasks

Dreamer’s stated platform goal is to create more value for participants than for itself. Tool builders already receive payments; monetization for shared agents remained under consideration. The announced Builders-in-Residence program offered expedited access, paid places for selected creative builders, and direct support for missing capabilities. Inspired by artist residencies and Lego’s Master Builder program, it sought ideas beyond the company’s existing team and ecosystem.

A $10,000 tool-building prize was announced alongside a mid-April reference, without precise deadline or eligibility rules. Prospective builders were directed to Dreamer’s builder information for the programs and expedited waitlist access.

Commerce has two roles here: helping builders earn money on the platform, and letting agents complete transactions in the world. Dreamer reports that some tool builders already move money to finish tasks. Its strategy is to support competing commerce protocols, then invest more deeply if one or two dominate—an approach compared in the conversation to the early web’s competing protocols. A separate concern raised is that usage-based AI pricing encourages proprietary credits that obscure margins; their eventual consequences remain unresolved.

The next demo makes that boundary concrete. A Big Sky ski-trip app combines Ski Bum’s live lift status with a forecast, accommodation details, and a friend’s dining recommendations. Friends who were not yet on Dreamer received an iMessage link and used the app on their phones. They entered expenses and calculated who owed whom. It reconciled expenses but did not transfer money; payment execution was proposed as a tool-building opportunity.

Attain Finance supplies financial data for personal-finance agents through a system described as similar to Plaid. Dreamer says it conducted a detailed security review with the company before admitting the community-built tool. Bank of America connectivity was suggested with confidence, but not confirmed in the exchange.

Stable tools over changing models

Dreamer continually evaluates multiple leading models so users do not have to choose them. Its curated tools cover image understanding, image generation, transcription, text-to-speech, and feed exploration. Cooking requests prompted a recipes tool backed by a selected repository.

The intended contract is a stable interface with an evolving implementation. An app can keep calling the translation tool while Dreamer updates the model underneath it; users reportedly use it for long Chinese podcasts translated into English. Branded integrations remain available when the provider matters, such as Instacart for groceries. Building this catalog requires both dedicated partnerships work and outside contributions.

One participant calls this an agent lab: a company specializing in model evaluation, routing, and domain judgment rather than training models. The same participant identifies a tension: as multimodal capabilities merge, fixed categories such as image-understanding endpoints may become restrictive. Dreamer’s response emphasizes adaptive sub-agents that can change their approach based on the data.

Reuse also works above the integration layer: agents appear to the system as tools. Builders can package a preferred behavior once and invoke it from subsequent apps. The open tool layer can also expose external capabilities—including, hypothetically, a builder’s own foundation model.

The engineering surface behind the conversation

Agent Studio’s debug drawer exposes build logs, generated files, editable prompts, and file uploads. Dreamer built its software development kit (SDK) and command-line interface (CLI) before its conversational interface. Engineers can work locally with Cursor or Claude Code, or write code directly without Sidekick.

Sidekick uses the same CLI as human developers, accessing the platform under the user’s identity. Complete command coverage and documentation therefore serve both audiences: they explain available capabilities to agents while making the platform easier for engineers to use.

Applications run full code in a general-purpose VM. Other languages are possible through the SDK, but generated experiences default to TypeScript. Compiler feedback gives the coding agent a concrete repair signal: generate code, inspect type errors, and fix them. The team reports type safety across most of its own stack, from database to frontend, and carries that approach into the agent SDK.

Dreamer built many development tools internally, including LLM-driven code review. Application version history is custom rather than Git-based, with efficiency and understanding version contents offered as reasons. But the choice also reflected early expediency: it made versions easy for Sidekick to manipulate. The company uses Git for platform source and might choose it for this feature if starting again.

Agents can run on external triggers, including webhooks and incoming emails selected by filters, sometimes using an LLM. The activity view shows each run’s trigger, timing, and actions. Opening the conference planner’s Guide Me event reveals the individual model calls and actual prompts for each time slot. That task used Haiku, described as fast and high quality without quantitative measurements.

Identity and data sharing are platform responsibilities

Each agent gets an on-demand SQLite database, with row ownership handled for multiple users. Platform authentication carries a user’s identity into apps published by other builders. Queries return that user’s data unless the builder explicitly designates shared data. For the Big Sky app, the builder asked Sidekick to make everyone’s expenses visible to the group; user data would otherwise have been isolated by default.

Builders can inspect the app’s function breakdown, distinguish exported functions from internal ones, and invoke them individually. The database-manipulation technology is in-house, while underlying infrastructure uses outside providers. The team describes moving away from building so much infrastructure itself to focus on its distinctive product work.

Configurability has a deliberate boundary. Developers can use external coding tools through the SDK, but replacing the coding engine inside Sidekick was not supported. Dreamer prioritizes removing software-building complexity for people who could not previously implement their ideas; choosing an internal coding agent was not a priority for that audience.

Memory carries context between apps

Sidekick accumulates information about the user that other agents can draw on, showing which facts they consider relevant to a use case. A Weekend Activity Planner reportedly recommended a St. Patrick’s Day parade at Civic Center because its profile included the user’s Irish background, learned through Sidekick.

The early memory system stored facts in a vector database and retrieved them through embeddings. The team says it worked but was more complex than necessary, so it replaced it. Knowledge graphs were also tried; the system in use at the time was not a knowledge graph. Its replacement architecture was not specified, and quality claims remained qualitative. More than one person was dedicated to continued memory work.

A small team that hires for agent fluency

The company discussion frames the move from chief technology officer to chief executive as broader responsibility, greater personal accountability, and more time spent on reaching customers and talking directly with users. The argument for small teams centers on communication overhead: each new colleague adds relationships and coordination, making larger organizations harder to keep nimble.

A core group of about six people reportedly built the demonstrated product; the company had grown to about 17 by the conversation. Dreamer started in late 2024 and uses coding agents and its own marketing agents internally. Its hiring principle is mutual contribution: each new colleague should learn from the team and help existing colleagues improve.

A return to community software gives Attain Finance a memorable example. Its builder, an external early-alpha participant identified as Andrew, demonstrated an agent that analyzed purchases and recommended buying fewer burritos. This was personalized financial feedback, with no measured savings reported.

Working well with coding agents is a hiring criterion. Many team members had previously managed engineers but now worked as individual contributors. Management experience is presented as useful preparation for directing agents, provided the engineer remains close enough to the code to write and assess it.

Interviews begin with a coding screen, then move to short collaborative projects with a team engineer. Candidates discuss a developed product idea and build it using their preferred coding agent. Interviewers observe prompting, product judgment, and how candidates use time while agents run. Workflows include rotating between concurrent agents and having one agent review another’s code; the emphasis is on directing and inspecting work, without a measured productivity claim.

Taste remains the harder target

The closing discussion acknowledges subjective improvements in Suno music generation and NanoBanana image generation, including infographics and painterly images. The remaining weakness is taste: understanding people and their motivations well enough to create software with individuality and delight.

Dreamer’s team says it invested substantial design work in templates, prompts, and the agent harness to make generated experiences feel less generic. Producing a functioning app does not automatically produce a distinctive experience; that judgment still requires deliberate product work.

The other participant agrees that taste matters but questions how to make it a tractable research target. Recognizing a good result on sight offers no clear, verifiable evaluation criterion. Code generation supplies more concrete feedback; quantifying taste may require a deeper understanding of the creative process. The exchange ends on the gap between increasingly satisfying machine output and the remaining results people still want to improve.

The closing invitation is to try Dreamer and build. The recording offered listeners a limited-time waitlist bypass through Dreamer’s Latent Space page, without specifying an expiration date.