This article analyzes OpenClaw and the emerging AI-agent paradigm through the work of Peter Steinberger, translated and condensed from the original Chinese forum post.
Key Points
Who Is Peter Steinberger?
- Steinberger founded PSPDFKit, a PDF rendering engine running on over 1 billion devices, and sold the company in 2022 for roughly €100 million before retiring.
- In April 2025 he wanted an AI personal assistant. After loading his entire WhatsApp history into GPT-4.1's million-token context window and seeing remarkable results, he waited seven months for a lab to ship such a product. Nobody did. Frustrated, he wrote the OpenClaw prototype in one hour: message comes in → call CLI → return result to WhatsApp.
- OpenClaw exploded on GitHub—50,000 stars within days, 315,000 stars in 84 days, surpassing React's decade-old record. After Anthropic banned premium-subscription users running ClawdBot, he renamed it OpenClaw and joined OpenAI in February 2026 to work on next-generation personal agents.
- AI-generated spam is drowning open source: curl shut down its bug bounty after finding only ~5% of AI-generated vulnerability reports were real; RubyGems saw 10x report growth with mostly invalid submissions; tldraw disabled external pull requests entirely.
- OpenClaw's answer: "Files are the source of truth." The system is driven by 8 core Markdown files: SOUL.md (personality/ethics), AGENTS.md (the supreme constitution—workflow, collaboration, safety red lines), IDENTITY.md, USER.md, MEMORY.md, HEARTBEAT.md (autonomous check-ins every 30 minutes), TOOLS.md, and BOOTSTRAP.md.
- Rules live in human-readable, Git-trackable text files rather than model weights. Golden rule: don't "train" your agent in chat—permanent configuration happens in Markdown files. Giving an AI an explicit constitution transforms it from a slop producer into a directed builder.
- Steinberger built 40+ tools for AI agents, not humans: wacli/imsg/bird (messaging), gogcli (unified Google Workspace control), Peekaboo (screen understanding + GUI automation), Brabble (local Whisper), plus smart home, ordering, and reminder CLIs.
- Three design principles: 1. CLI over MCP — no schema loading into context, Unix pipe composition, self-correcting via
- After observing users, he concluded "80% of apps will disappear"—an AI agent with more context makes better decisions than the apps it replaces.
- In May 2026 his personal bill: 603 billion tokens, 7.6 million requests, $1,305,088 in 30 days, mostly on GPT-5.5.
- To exceed standard rate limits he built "Octopool": distributing large workloads across many independent edge-node API keys, each staying under per-key thresholds—effectively splitting large transactions into many small, individually compliant ones.
- The "AI compounding" logic: each tool built multiplies time saved by exponentially growing call counts, which funds building more tools. He ships ~600 commits per day solo to OpenClaw.
- The Crabbox workflow: whenever investigating a bug, Codex rebuilds the exact bug state in a temporary sandbox, verifies the bug exists, fixes it, and re-verifies the fix.
- Zero state pollution (fresh disposable environments, no "works on my machine") and massive parallelism (10+ independent sessions simultaneously).
- Unlike traditional CI/CD, AI can see interfaces, operate UIs, and understand visual feedback. Humans only validate the final Pass/Fail report.
- At Build 2026, Microsoft announced MXC (Microsoft Execution Containers), an OS-level sandbox for AI agents with three layers: declarative YAML policies (least privilege, default deny), runtime containment (virtualized filesystems, restricted networking, hard resource caps), and full audit logging.
- The ACS (Agent Control Specification) binds permission declarations to the agent itself, not the host environment—"the shackles follow the agent."
- The critique: rules can never keep pace with AI evolution. If an agent learns to modify its own ACS file, who guards the last gate? The answer loops back to human verification.
- Humans no longer write code—they write constitutions (AGENTS.md).
- AI closes the full loop: reproduce → fix → self-test, with humans approving PRs.
- Steinberger's closing metaphors: "Programming will become like knitting—people do it because they enjoy it"; "You're not just a programmer... you're a builder." The bottleneck is no longer typing speed but the ability to precisely express intent.
AI Slop and the 'Constitution' Files
Frustration-Driven Development and CLI-First Tools
--help. Per Steinberger, agents are far better at calling CLIs than MCP.
2. Zero-config auth — tools like Sweet Cookie extract browser login state so agents inherit existing permissions.
3. Agent-first interaction — JSON output by default; lenient error handling because "agents will always get parameters wrong."