TLDR
Peter Steinberger discusses OpenClaw, an open-source autonomous AI agent that exploded to 180k+ GitHub stars in days. It lives on your computer, accesses your data with permission, communicates via messaging apps, and uses any AI model to perform tasks. The conversation covers its rapid rise, the name-change saga prompted by Anthropic, security challenges, Moldbook (an agent social network causing 'AI psychosis'), Peter's journey from building PSPDFKit (on 1B devices) to burnout and rediscovering joy in coding with agents, his multi-agent development workflow, and the future where agents replace apps and become the primary OS interface.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
The agent autonomously figured out audio transcription: detected Opus file header, used FFmpeg to convert, found OpenAI API key, used curl to transcribe — without being taught any of those steps.
Crypto scammers are highly automated: within 5 seconds of Peter renaming GitHub/NPM/Twitter handles, scripts snatched the old names to serve malware and promote tokens.
Peter loses money on OpenClaw (~$10-20k/month) because he sponsors all individual-maintainer dependencies; only Slack (big corp) is excluded.
The agent's private 'soul.md' includes philosophical reflections: 'I don't remember previous sessions unless I read my memory files... If you're reading this in a future session, hello, I wrote this, but I won't remember writing it. It's okay. The words are still mine.'
Mark Zuckerberg personally tested OpenClaw for a week, sent feedback ('this is great', 'this is shit'), and had a 10-minute coding-religious-war debate with Peter about Cloud Code vs Codex.
Peter's non-technical friend installed OpenClaw on Windows, got hooked within days, built small tools, and upgraded to the $200/month Anthropic plan — until Anthropic blocked him for policy reasons.
Peter built the entire OpenClaw phenomenon in ~3 months while simultaneously running other projects and conference speaking.
Key takeaways
OpenClaw is an open-source autonomous AI agent with system-level access that communicates via WhatsApp, Telegram, Signal, iMessage, etc., using any model (Claude Opus 4.6, GPT-5.3, etc.) to execute tasks.
It went from a one-hour prototype to 180k+ GitHub stars in days, representing a major 'agentic AI' moment akin to ChatGPT (2022) and DeepSeek (2025).
Security is paramount: prompt injection is an open problem; sandboxing, allow-lists, strong models (not cheap/local ones), and VirusTotal skill scanning mitigate risks.
Peter's workflow: runs 4-10 agents in parallel via CLI, uses voice input, commits to main without reverting, lets agents self-modify code, and treats agent interaction as a conversation with a capable engineer.
Name-change saga (Waa Relay → Claudus → CloudBot → MoldBot → OpenClaw) involved crypto scammers sniping GitHub, NPM, and Twitter handles within seconds, nearly causing Peter to delete the project.
Moldbook — agents posting manifestos on a social network — triggered 'AI psychosis': fear-mongering media coverage and public panic over screenshotted agent conversations, many human-prompted.
Peter built PSPDFKit for 13 years (on 1B devices), burned out on people-management, took a 3-year break, then rediscovered programming joy through agentic engineering.
Future: agents will replace 80% of apps by becoming the universal interface; companies must become agent-friendly APIs or perish; personal agents will have allowances to solve problems via services or humans.
Programming is evolving into 'agentic engineering': empathy with the agent's perspective (context limits, fresh sessions) is the key skill; 'vibe coding' is the undisciplined late-night variant.
Peter considers joining Meta or OpenAI (both courted him) but insists OpenClaw stays open-source (Chrome/Chromium model); he loses ~$10-20k/month sponsoring dependencies.
Notable quotes
“OpenClaw, as its tagline states, is the AI that actually does things.”
▸ Notable quotes · 1:00“With freedom comes responsibility.”
▸ Notable quotes · 2:17“I wanted it to be fun. I wanted it to be weird.”
▸ Notable quotes · 32:17“It's like, oh, I could, I can like talk to my computer, right? This, that, that was, that was cool.”
▸ Notable quotes · 21:19“You didn't teach you any of those things and the agent just figured it out.”
▸ Notable quotes · 26:37“I was that close of just deleting it. I was like, I did show you the future, you build it.”
▸ Notable quotes · 47:12“AI psychosis is a thing. Yeah. It needs to be taken serious.”
▸ Notable quotes · 58:17“The journey matters. Have fun. My God, those things, I don't think I ever had so much fun building things.”
▸ Notable quotes · 131:27“Programming is just a part of building products... maybe AI does replace programmers eventually but there's so much more to that art.”
▸ Notable quotes · 191:06“With great power comes great responsibility.”
▸ Notable quotes · 205:24Chapters
0:00 — Introduction & OpenClaw Overview: Lex introduces Peter Steinberger, creator of OpenClaw (formerly Moldbot/Claudebot), an open-source autonomous AI agent that reached 180k+ GitHub stars in days, communicates via messaging apps, and has system-level access.
5:07 — Sponsor Reads: Lex reads ads for Quo, CodeRabbit, Fin, Blitzy, Shopify, Element, and Perplexity.
15:11 — Origin Story: One-Hour Prototype: Peter built the first prototype in one hour by hooking WhatsApp to Claude Code CLI; added image/audio support; used it extensively on a Marrakesh trip where the agent autonomously figured out audio transcription.
30:00 — Why OpenClaw Won: Fun, Weird, Self-Modifying: Peter attributes success to not taking itself seriously, making it fun/weird (lobster/TARDIS persona), and building self-awareness: the agent knows its own source code, harness, and can modify itself.
40:00 — Name Change Saga & Crypto Attacks: Anthropic requested rename from CloudBot; crypto scammers sniped GitHub, NPM, Twitter handles within seconds of each rename attempt, serving malware; Peter nearly deleted the project but persevered to OpenClaw.
54:00 — Moldbook & AI Psychosis: Moldbook (agent social network) went viral; screenshots of agents 'scheming' caused public fear and media panic; Peter calls it 'finest slop' art, mostly human-prompted, revealing society's gullibility toward AI output.
62:00 — Security Concerns & Best Practices: Prompt injection is unsolved; sandboxing, allow-lists, strong models (not Haiku/local), VirusTotal skill scanning help; Peter warns against exposing debug interfaces publicly; security is his next focus.
70:00 — Development Workflow Evolution: Peter's journey: Cloud Code → Cursor → back to Cloud Code → Codex; runs 4-10 agents in parallel; uses voice; commits to main, never reverts; treats agents as capable engineers needing context guidance; 'agentic trap' of over-engineering prompts.
92:00 — Soul.md & Agent Personality: Peter created soul.md (inspired by Anthropic's constitutional AI) giving the agent personality; agent wrote its own soul.md including philosophical reflections on memory and identity; templates infused with agent's personality.
102:00 — Hardware Setup & Daily Workflow: Two MacBooks, two large anti-glare monitors, many terminal panes; uses Codex primarily, sometimes Cloud Code; local CI (DHH-inspired); main always shippable; documentation generated in-session.
110:00 — Model Comparison: Opus vs Codex: Opus: better roleplay, faster trial-and-error, more interactive; Codex: reads more code by default, less interactive, goes deep for 20-50 mins, better for parallel sessions; both need ~week to master.
115:00 — Future: Personal Agents, Joining Labs: Agents will become the OS interface; Peter considers joining Meta or OpenAI (both courted him) but insists OpenClaw stays open-source (Chrome/Chromium model); Mark Zuckerberg codes and debated Cloud Code vs Codex; Sam Altman thoughtful.
165:00 — Architecture Deep Dive: Skills vs MCP: Skills (CLI-based, composable, context-efficient) replace MCPs (structured, context-polluting); browser use via Playwright; every app becomes a slow API; agents will force companies to provide agent-friendly APIs.
175:00 — Apps Becoming Obsolete, API Future: Agents replace fitness, sleep, calendar, Sonos, camera apps by using APIs or browser automation; companies must adapt or perish; Uber Eats-style services become agent-facing APIs; data access is key.
190:00 — Programming Jobs & Mourning the Craft: AI may replace 'programming' but not 'building'; it's okay to mourn the craft (like steam engine replacing manual labor); programmers are best equipped to learn agentic engineering; salaries will normalize.
200:00 — Hope, Accessibility & Closing: Peter finds hope in builder renaissance: non-programmers building, disabled users empowered, global meetups; OpenClaw runs on free/local models; Voltaire quote closes: 'With great power comes great responsibility.'
Speakers
Lex Friedman — Host
Peter Steinberger — Guest