How Peter Steinberger Built a Viral AI Agent Solo

Peter Steinberger — creator of PSPDFKit (1B+ devices) turned OpenAI agent lead — sits down for the inaugural episode of OpenAI’s “Builders Unscripted” series. Having recently joined OpenAI after his open-source personal AI agent OpenClaw amassed 180K+ GitHub stars and spawned a community event (ClawCon) drawing over a thousand attendees, Steinberger shares the deeply personal journey from burnout to building what he calls “the final form” of what he always wanted to create.

On the moment AI clicked: After 13 years running PSPDFKit, Steinberger burned out and stepped away from tech. When he returned, he fed a half-finished project into Gemini Studio to generate a spec, then dragged that spec into Claude Code and typed “build.” “One hour later, it actually worked. It was the worst slop. But this was the moment where it really clicked — I got goosebumps for the possibilities.” That dopamine hit led to a year of obsessive building.

On agents as problem solvers, not just coders: When Steinberger sent a voice message through his WhatsApp bot, the agent figured out how to handle it entirely on its own — inspecting the file header, converting with FFmpeg, finding an OpenAI key in the environment, and using cURL to transcribe via the API. “We built these things for agentic engineering. But really, the skill is more abstract — if you want to be a really good coder, you need to be a really good problem solver. And that just maps in any domain.”

On the “agentic trap”: Steinberger warns that many developers get stuck over-optimizing their AI setup instead of actually building. “A lot of people get stuck trying to super optimize their setup. It doesn’t really make you more productive, but it feels like you’re more productive.” His own approach is deliberately simple: no worktrees, just 1-10 checkouts, and conversational prompting where he always asks the model “Do you have any questions?”

On vibe coding vs. agentic engineering: In a signature take, Steinberger doubles down on his controversial position: “I think vibe coding is a slur. They try AI, but they don’t understand that it’s a skill. You pick up the guitar — you’re not going to be good at the guitar in the first day.” The difference is treating AI coding as a craft that requires practice, not a magic button.

On PRs becoming “prompt requests”: With 2,000+ open PRs on OpenClaw, Steinberger has shifted how he reviews contributions. He doesn’t read the code first — he asks the model “What’s the intent of this PR?” Because many contributors used AI to generate their code, the intent matters more than the implementation. Often, a localized fix reveals a deeper architectural issue that needs a different solution entirely.

6 Takeaways from Steinberger on Solo AI Development

• One person can build what teams couldn’t — 90,000 GitHub contributions across 120+ projects in a year, something that “would not have been possible by any one human” before AI agents
• Optimize your codebase for agents, not humans — The code structure that helps AI work best isn’t always what a human developer would prefer, and that’s okay
• Approach it playfully — Don’t start with work projects; build something you’ve always wanted to build. The learning happens through exploration
• Trust builds over time — Steinberger’s trust in Codex “building what I want, from all the tools right now, is the highest,” but this came from months of practice
• The agent will surprise you — When placed in a Docker container with no cURL, the agent built its own HTTP client from a C compiler and TCP sockets
• Self-modifying software is here — OpenClaw users can literally prompt the agent to change its own source code, which is why non-developers are submitting PRs for the first time

What This Means for Solo Builders and AI Agent Adoption

Steinberger’s story is a proof point for the “one-person unicorn” thesis that Sam Altman has championed. The combination of deep domain expertise, agentic tooling, and a builder’s mindset creates leverage that simply didn’t exist a year ago. His advice to European developers who haven’t adopted agentic tools yet: “If you are high-agency, if you’re smart, you will be in more demand than ever.” The question isn’t whether to adopt AI agents — it’s how quickly you can develop the skill to wield them effectively.