Peter Steinberger: “Fun Is Velocity” — an OpenClaw Retrospective (#AI)
I watched a talk by Peter Steinberger, the creator of OpenClaw, at Y Combinator Startup School 2026. I have already covered his interview about local-first agents and the claim that “80% of apps will disappear.” This is a different format: an inside look back at eight months of the project—what went wrong, why its creator stopped using his own product, and why “fun is velocity” is a measurable signal rather than a slogan.
For context, OpenClaw began in November 2025 with a personal frustration: Peter wanted to talk to his coding agents from his phone, so he built a WhatsApp relay to an agent running on his machine. The project went viral after New Year. According to Peter, at its peak it reached 4,7 million downloads per week and had around 3000 people contributing commits to the repository. I have already posted a joke about what it is like to set up such a self-hosted agent :)
The most valuable part of the talk is its candid account of the mistakes.
1️⃣ Feature bloat as a systemic effect of open source Every community feature needed a configuration option so that existing installations would not break; at the peak, the project had around 9500 options. That leads to the talk’s best quote: “In the weeks when I enjoyed building, the product improved noticeably. In the weeks when I didn’t, we shipped configuration options.”
2️⃣ “Your dependency’s business model is your business model” OpenClaw was optimized for subscription access to Claude Opus, and Anthropic’s decision to close that access made the pivot painful. Amusingly, the CEO of OpenCode described the same January event as an accelerator of his company’s growth. I covered his account: two products experienced the same dependency shock in opposite ways.
3️⃣ Security and media panic Peter says the media reported that “20% of skills were malicious,” while his team’s own analysis found 0,3% among 67 thousand skills. I am repeating both figures as his claims, without independent verification. His broader point is that “a correction never travels as far as the scare story.” Security professionals are skilled at producing scare stories, and often the threat model itself matters less to them than creating noise.
4️⃣ Losing himself as the first user Peter had started building “for everyone” rather than for himself and became the person who fixed bugs and triaged security reports. By May, downloads had fallen to 835 thousand per week—and recovered once he returned to features he personally needed.
From an engineering perspective, OpenClaw is most interesting as an experiment with an agent that lives neither in an IDE nor in a vendor’s cloud, but on your own machine: shell, files, and browser, with an ordinary messenger as the interface. The talk also exposes the limits of that architecture: the economics of an always-on agent whose idle heartbeat loops burn tokens—exactly the subject of my post on token burn—plus distributing load across machines and establishing trust in the skills ecosystem. It is essentially a map of the traps awaiting anyone building personal or enterprise agents.
Peter’s conclusions: you should be your product’s first user; fix what personally annoys you; when building becomes cheap, attention becomes expensive, so focus wins; and “everything you build can be forked—the only thing that cannot be forked is your name.”
My own takeaway is that “fun is velocity” is not about entertainment. It is a sign that you are building from the position of a user and have retained your taste. Once the product turns into an obligation, speed falls together with decision quality; in Peter’s case, the download graph makes that visible.
#AI #Agents #Product #Engineering #Management #Startup