Peter Steinberger¶
Austrian engineer; creator/BDFL of openclaw — the personal open-source AI agent project that became the fastest-growing project in GitHub history in early 2026. Previously known for PSPDFKit. Joined openai in 2026 to work on "bringing agents to everyone"; simultaneously runs the openclaw-foundation.
Persona¶
- European-at-heart data-ownership ethos ("you want to own your data") — motivated him to build a local-first personal agent rather than a hosted SaaS.
- "Madness with a touch of science fiction" — his self-description of how he runs projects. Frames OpenClaw as "only possible because it came out of Europe — legal would have killed this in any American company."
- Explicit hacker-energy disposition: "What's the worst that can happen? It could exfiltrate my tokens, nothing that would completely kill me. Worst pics are already online if you use Grindr."
- Runs 5–10 parallel coding-agent sessions simultaneously (Codex GPT-5.1, formerly more before speed improvements).
Workflow signature¶
- Prompt-request over pull-request framing — agent workflows as his default medium.
- Heavy on SOUL.md personality file — open-sourced his approach; turned the agent into a WhatsApp-native character that texts like a human (not too wordy, fewer dots, lobster jokes).
- Ships features by trying them on himself first; rejects the pure "dark factory" model because "the way to the mountain is usually never a straight line."
Key claims¶
- Taste is the new engineering moat. Lowest bar: "you immediately know if something stinks like AI." Higher: delightful details agents won't produce on their own (e.g. OpenClaw's occasional roast messages).
- System design still matters — "everything is in the clanker but you still need to ask the right questions, otherwise that makes the difference of good code vs slop."
- Saying no is a skill — every wild idea is a prompt away; the bottleneck is big-picture synthesis, not output.
- On security theater: "The higher they scream how critical the advisory is, the more likely it's slop."
Cross-references¶
- openclaw — the project
- alex-krentsel — Berkeley PhD who did the systems-side architectural deep-dive on the project Peter built
- openclaw-foundation — governance vehicle ("Switzerland for contributors")
- andrej-karpathy — runs OpenClaw at home; llm-knowledge-bases parallels OpenClaw's wiki/memory direction
- eric-zakariasson / mitchell-hashimoto — parallel voices on parallel-agent workflow
- agent-taste — concept he centers
- agent-security-slop — his framing of AI-generated security-advisory flood