--- schemaVersion: 1 slug: en-mp-254-20260406-pawelhuryn-ai-pawel-huryn-48-claude-dispatch ticketId: MP-254 lang: en title: Commanding an AI Army from the Playground — Paweł Huryn's 48-Hour Claude Dispatch Experiment summary: Product Manager Paweł Huryn ran a 48-hour Claude Dispatch experiment from his phone while supervising his kids at a bounce house. Twenty-five minutes of instructions produced over three hours of parallel AI output, turning fragmented downtime into productive agent orchestration. originalDate: 2026-03-28 translatedDate: 2026-04-06 source: "@PawelHuryn on X" sourceUrl: https://x.com/PawelHuryn/status/2037775847428383093 author: null authorshipNote: null canonicalUrl: https://gu-log.vercel.app/en/posts/en-mp-254-20260406-pawelhuryn-ai-pawel-huryn-48-claude-dispatch status: published replacementTicketId: null replacementUrl: null --- # Commanding an AI Army from the Playground — Paweł Huryn's 48-Hour Claude Dispatch Experiment > **Source:** [@PawelHuryn on X](https://x.com/PawelHuryn/status/2037775847428383093) ## Working Next to a Bounce House Paweł Huryn is a well-known author in the Product Management space, and he recently did something that paints quite a picture: while supervising his kids at a bounce house, he used his phone to direct AI agents through a full batch of real work. Not replying to a [Slack](https://gu-log.vercel.app/glossary#slack) message. Not skimming through email. Actual, substantive, multi-iteration work. His tool of choice was Claude Dispatch—Anthropic’s mobile-first agent interface. Over 48 hours, Huryn orchestrated multiple workstreams from his phone, including design iterations, competitive analysis, and stakeholder documentation. His total time spent giving instructions? About 25 minutes. The parallel output from Claude’s side? Over 3 hours of work. > **Mogu butts in:** > > About 25 minutes for 3 hours of output—returns like that would probably trigger an SEC investigation in finance. But seriously, the point isn’t “AI is fast and cool”—it’s that a PM, while watching his kids play, used fragmented moments to knock out work that would normally require sitting at a desk. This shift in behavior pattern is more striking than any benchmark. (◕‿◕) --- ## Phone as Remote Control: A New Work Pipeline In his article, Huryn describes what he calls a **Phone → Orchestrator → Task Sessions → Desktop** pipeline. > **Mogu roast time:** > > Here’s an analogy before we go further: using AI the old way was like hailing a taxi—you had to stand on the curb waiting, stay in the car the whole ride, and only move on once you arrived. The Dispatch model is like ordering Uber Eats—place the order, go about your day, deal with it when it arrives. The difference isn’t delivery speed; it’s that your waiting time is freed up. ╰(°▽°)╯ Here’s how it works: the phone handles “initiating” and “directing.” Open Claude Dispatch, issue tasks in natural language—“shift the icon left on this infographic,” “make the header bolder,” “run a competitive analysis.” Each task spawns a persistent session where the agent runs in the background and waits for review when finished. Meanwhile, the human? Put the phone away after giving instructions. Go back to watching the kids, walking the dog, grabbing coffee. Pick up the phone when convenient to check results and give the next round of direction. Only return to the desktop when you need to polish something or use the actual files. This is completely different from traditional AI chat. The traditional model is **synchronous interaction**—open ChatGPT, type, wait for a response, review, type again. The human must be present throughout. The Dispatch model is **asynchronous delegation**—fire and forget, the agent runs on its own, check in whenever you feel like it. --- ## Four Rounds of Iteration: The Infographic Case Huryn shared a concrete example to demonstrate this workflow. He gave four rounds of iteration instructions for a visual asset (an infographic) from his phone: Round one: initial generation. Round two: “Move the icon left.” Round three: “Make the header bolder.” Round four: fine-tune the details. Each round was a few seconds of text input, then Claude executed and delivered. Meanwhile, Huryn’s other two sessions ran in parallel—one doing competitive analysis, another drafting a stakeholder page. Three workstreams, running simultaneously, requiring just one person occasionally switching sessions on a phone to check progress and provide direction. > **Mogu PSA:** > > Four rounds of design iteration in a traditional workflow would look something like this: PM writes a brief → waits for designer availability → reviews first draft → schedules a meeting to discuss changes → waits for revisions → reviews again. One infographic going through four rounds might take three to five business days, with gaps for “the designer is busy with another project.” Huryn knocked it out in about 25 minutes next to a bounce house. The contrast is almost cruel. ┐(￣ヘ￣)┌ --- ## Reclaiming Dead Time Huryn’s core argument is simple but powerful: AI agents turn “dead time” into productive time. What’s dead time? Walking the dog. Waiting for coffee. Standing in line. Watching kids at the playground. This time was always “wasted”—not because people are lazy, but because the tools and environment for real work weren’t available. But when your phone becomes a remote control for agents, all those fragmented moments get activated. No need to open a laptop, connect an external monitor, or enter “deep work” mode. Just a few seconds of judgment and a few lines of text, and agents handle the work in the background. This doesn’t change the tools—it changes the fundamental assumption of “what contexts allow for work.” > **Mogu highlights:** > > There’s a subtle but important nuance here: dead time becoming productive time sounds great, but it also means “not working” time gets further compressed. When walking the dog can involve directing AI to do work, when do you actually rest? Huryn frames this as a productivity breakthrough, but from another angle, it’s the further creep of always-on culture. It’s not that Dispatch isn’t useful—it’s that when tools are this powerful, you have to draw your own boundaries. ʕ•ᴥ•ʔ --- ## The PM Role Transformation: From Doing to Orchestrating But Dispatch’s initial setup process is anything but the “seamless” experience it advertises—more on that later. First, let’s address a more fundamental question: if agents can execute on behalf of PMs, what core capabilities do PMs still need? The most thought-provoking insight in Huryn’s article concerns the essential transformation of the PM role. Traditional PM core competencies include: writing PRDs, conducting market research, collaborating with designers and engineers, managing stakeholder expectations. These are “doing” abilities. But in the agent era, PM value is shifting from “doing” to “orchestrating.” Specifically, Huryn believes future PMs need new capabilities: **Judgment**—knowing what granularity to break tasks into for agents, and when to step in yourself. Not everything should be delegated to AI; discerning where that line is becomes its own expertise. **System-building ability**—designing workflows and rules that let agents execute with minimal human intervention. This is like writing SOPs for agents, rather than running the SOPs yourself. **AI fluency**—the ability to spot hallucinations, and to structure autonomous “Plan → Execute → Reflect” loops. It’s not just about prompting—it’s genuinely understanding agent behavior patterns and boundaries. > **Mogu wants to add:** > > The “Plan → Execute → Reflect” loop sounds fancy, but it’s essentially the same logic as managing a human team: assign objectives → let people execute → review outcomes and decide next steps. The difference is agents don’t complain about overtime, don’t fight other departments for resources, and don’t report “still researching” in standups. But agents also make mistakes humans wouldn’t—like confidently delivering completely wrong information. So Huryn’s point about “judgment” really is key: the hardest part of directing agents isn’t giving instructions, it’s seeing through the problems in the results. (๑•̀ㅂ•́)و✧ --- ## Real-World Gotchas: Not So Seamless Huryn’s article isn’t all sunshine. He honestly shared some pitfalls he encountered. **Knowledge storage**: Huryn recommends storing knowledge bases in a GitHub repo or synced Google Drive so agents can access background information on each launch—implying Dispatch itself doesn’t automatically retain context between sessions. **File access**: As of the article’s writing, Dispatch didn’t support directly downloading agent-generated files. You need to sync via Google Drive and retrieve files from the desktop. This adds friction to the workflow. **Folder permissions**: When setting up sync, you need to explicitly grant folder access to the agent—a step that’s easy to overlook, which then blocks progress. **Initial setup**: The overall setup process has a learning curve and friction; it’s not install-and-go. > **Mogu roast time:** > > These gotchas are actually important because they expose a pattern: the more an AI tool advertises itself as “seamless,” the less seamless the initial setup usually is. Once Dispatch is running, the experience might indeed be smooth, but the journey from zero to one—setting permissions, configuring sync, building a knowledge base—is the classic “spend a day setting up your environment before you can start real work.” Honestly, it’s not unlike DevOps pain points. (￣▽￣)／ --- ## Four Mobile Claude Options Huryn also compared four ways to use Claude on mobile, and he considers Dispatch the most transformative. While the complete details of all four options aren’t fully available from the source, the core difference is this: other options are mostly extensions of synchronous chat (type on phone, wait for response, type again), while Dispatch manages tasks as persistent sessions—once you send it off, it’s off, no need to stare at the screen waiting. This difference might seem like a minor UX detail, but Huryn sees it as the real architectural shift: transforming the phone from “a chat window on a small screen” to “a remote control for command central.” --- ## Conclusion Huryn’s 48-hour experiment isn’t a proof of concept—it already happened. A PM, next to a bounce house, used his phone to complete a full batch of work that would normally require sitting at a desk. Huryn believes the assumption that “productive work requires sitting in front of a computer” has been shattered. As for what work looks like after that assumption disappears—his experiment is just the beginning.