OpenClaw: When AI Fixes Its Own Bugs, Humans Are Left With Verification

Forum topic · QianXun · 2026-06-13

Summary

This long-form analysis of the OpenClaw project and its creator Peter Steinberger describes how AI agents are reshaping software development. Steinberger, former founder of PSPDFKit, built OpenClaw as a one-hour prototype after waiting seven months for an AI personal assistant that never arrived. The system is governed by eight Markdown 'constitution' files (SOUL.md, AGENTS.md, etc.) embodying the principle 'files are the source of truth,' contrasting with the flood of AI-generated slop overwhelming open-source projects like curl, RubyGems, and tldraw. Steinberger built 40+ CLI tools for agents rather than humans, rejecting MCP in favor of CLIs with JSON output and forgiving error handling. His 'Octopool' architecture reportedly pushed 40 million tokens per minute and consumed $1.3 million in API costs in 30 days. The 'Crabbox' workflow lets Codex agents reproduce, fix, and verify bugs in disposable sandboxes in parallel, reducing the human role to verifying results. The article also examines Microsoft's MXC (Microsoft Execution Containers), an OS-level declarative policy sandbox for agents, and questions whether such 'child locks' can keep pace with AI evolution. It concludes that programming power is shifting from 'doing' to 'validating,' and that future builders will orchestrate natural language, CLI toolchains, and AI agent fleets.

*Translation and summary of a viral Chinese tech forum post analyzing OpenClaw, Peter Steinberger, and the shifting role of human programmers in the AI-agent era.*

Key points

Origin story: Peter Steinberger, who founded PSPDFKit (a PDF engine running on 1B+ devices, sold for ~€100M in 2022), wanted an AI personal assistant in April 2025. After waiting seven months for anyone to build it, he wrote the OpenClaw prototype in one hour. It then exploded on GitHub, reportedly hitting 315,000 stars in 84 days.
Conflict with Anthropic: When Anthropic banned ClawdBot users on premium subscriptions, Peter renamed the project OpenClaw and joined OpenAI in February 2026, hired personally by Sam Altman to work on "next-generation personal agents."

AI slop vs. a "supreme constitution"

The post contrasts OpenClaw with the crisis of AI slop in open source: curl shut down its bug bounty after finding only ~5% of AI-generated reports were real; RubyGems saw 10x report growth with months of no valid submissions; tldraw disabled external PRs entirely.
OpenClaw's answer: "Files are the source of truth." The system runs on 8 core Markdown files:

| File | Role | |------|------| | SOUL.md | Personality, tone, ethics | | AGENTS.md | "Supreme constitution" — workflow, safety red lines | | IDENTITY.md | Daily role and skills | | USER.md | Owner profile and preferences | | MEMORY.md | Long-term, cross-session memory | | HEARTBEAT.md | Autonomous 30-minute self-inspection loop | | TOOLS.md | Local tool usage guide | | BOOTSTRAP.md | One-time init, destroyed afterward |

Rules are externalized into editable, Git-tracked text rather than baked into model weights. Prompt assembly: message → scan workspace → read all Markdown files → inject into system prompt → call LLM.

Frustration-Driven Development

Peter's 40+ tools (wacli, imsg, bird, gogcli, Peekaboo, Brabble, camsnap, sag, sonoscli, etc.) are built for agents, not humans, on three principles:

CLI over MCP

--help

Zero-config auth

Agent-first interaction

Implication: "80% of apps will disappear" because agents have more context and need no UI or subscriptions.

Token laundering and AI compounding

In May 2026, Peter reported 603 billion tokens, 7.6M requests, and $1,305,088 in API spend in 30 days (mostly GPT-5.5), hitting a rate limit of 40 million tokens/min until Sam Altman replied the limit was "too low."
"Octopool" shards requests across many edge-node API keys — described as "token laundering" to bypass rate limits.
AI compounding: each tool's savings scale with exponentially growing call counts; one person shipping 600 commits/day matches a traditional three-year team effort.

Crabbox: AI verifying AI

> "Whenever I investigate a bug, I have Codex recreate the exact bug state in a fresh Crabbox, verify the bug exists, fix it, and verify the fix."

Closed loop: reproduce → verify → fix → re-verify, each in a disposable sandbox with zero state pollution, 10 sessions in parallel.
Humans shift from "doing" to "accepting": only the final Pass/Fail report needs review.

MXC: Microsoft's "child lock"

At Build 2026, Microsoft announced MXC (Microsoft Execution Containers), an OS-level agent sandbox with three layers: declarative YAML policy (least privilege, default deny), runtime containment (containers, capped resources), and full audit logging.
ACS (Agent Control Specification) binds permissions to the agent itself, not the host environment.
The post's critique: rules are written slower than AI evolves — and if an agent can edit its own ACS file, who is the last gatekeeper?

Conclusion: the transfer of power

The five threads converge: humans no longer write code but write constitutions; high-bandwidth human attention is the only scarce resource; AI compound interest lets one builder do team-scale work; Crabbox closes the self-verification loop; MXC tries (imperfectly) to leash agents.
Peter's framing: "Programming will become like knitting — people do it because they love it, not because it's necessary." The bottleneck is no longer writing code but precisely expressing intent. Echoing Karpathy: "In five years, those who can't program in natural language will be like developers still insisting on assembly."

> "The future belongs to programmers who can tell stories." — Peter Steinberger

*The original post cites interviews with Peter Steinberger, OpenClaw's technical architecture, Microsoft Build 2026, and sources including Nature, arXiv, LeadDev, and The New Stack. Note: the events described are dated 2026 and reflect the original post's claims.*