State of the Claw — Peter Steinberger

Key Concepts

• Open Claw: An open-source AI agent framework, currently the fastest-growing project in GitHub history.
• Agentic Systems: AI systems capable of performing tasks, interacting with data, and communicating autonomously.
• Bus Factor: A measurement of the risk resulting from information and capabilities not being shared among team members (i.e., how many people could be "hit by a bus" before the project stalls).
• Security Advisories (CVEs/CVSS): Vulnerability reports; Open Claw has faced a high volume of these due to its rapid growth and "insecure by default" perception.
• Prompt Injection: A security vulnerability where an attacker manipulates an AI model's input to override its instructions.
• Dreaming: A conceptual framework for AI agents to reconcile memories, perform "garbage collection" on logs, and convert short-term interactions into long-term storage.
• Taste: The ability to discern high-quality, human-like output from "AI slop" (low-quality, generic, or poorly integrated AI-generated code/content).

1. State of the Project: Open Claw

Peter Steinberger, creator of Open Claw and current OpenAI employee, provided an update on the project’s trajectory:

• Growth Metrics: The project is five months old, boasting ~30,000 commits, nearly 2,000 contributors, and approaching 30,000 Pull Requests (PRs).
• Organizational Structure: Steinberger balances two roles: working at OpenAI and running the Open Claw Foundation. The foundation is being established to ensure the project remains neutral ("Switzerland") and sustainable, independent of any single corporate entity.
• Collaborations: The project has gained support from major industry players, including Nvidia (hardening the codebase), Microsoft, Red Hat, Tencent, and ByteDance.

2. Security Challenges and Real-World Applications

Open Claw has been a primary target for security researchers, leading to a high volume of vulnerability reports.

• The "Insecure" Narrative: Steinberger argues that much of the "insecurity" is a result of users ignoring recommended configurations (e.g., running agents in sudo mode or in public group chats).
• Data & Statistics: The project has received 1,142 security advisories (approx. 16.6 per day), with 99 classified as critical. 60% have been closed.
• Case Study (Nvidia Nemo Claw): Nvidia launched a security layer for Open Claw. Steinberger noted that when hooked to advanced models, it identified five ways to break the sandbox in 30 minutes, highlighting that more powerful models are better at both attacking and defending systems.
• Supply Chain Risks: The project has been affected by third-party dependencies (e.g., issues with Axios via Slack/Teams integrations), illustrating the complexity of modern software supply chains.

3. Methodology: The "Clanker" Workflow

Steinberger discussed his personal development philosophy:

• Iterative Development: He rejects the "waterfall" model. He advocates for building in steps, playing with the system, and allowing the project to evolve organically rather than sticking to a rigid initial plan.
• Token Maxing: He manages multiple agent sessions simultaneously (often 5–6 windows) to maintain high velocity, viewing this as a temporary workaround until model latency decreases.
• The "Dreaming" Concept: Inspired by human sleep cycles, this framework aims to help agents process daily logs, consolidate memories, and perform "garbage collection" to improve long-term performance.

4. Key Arguments and Perspectives

• The Importance of "Taste": Steinberger emphasizes that "taste" is the new moat for engineers. It involves recognizing the "smell" of AI-generated slop and ensuring UI/UX details feel human and intentional.
• Open Source vs. Corporate Control: He argues that Open Claw could not have been built within a traditional American corporate environment due to legal and risk-aversion constraints. Its success relies on its open, "hacker-first" nature.
• The Role of AI at Work: He believes that by allowing individuals to use AI at home, they will naturally demand similar, managed, and secure AI tools in their professional environments, driving enterprise adoption.

5. Notable Quotes

• "The higher they scream how critical they are, the more likely it's slop." — Steinberger on the nature of some security reports.
• "The way to the mountain is usually never a straight line... you find shortcuts, you see something new that inspires you." — On his iterative development philosophy.
• "If you don't think about [system design], you will eventually swipe yourself into a corner." — On the necessity of architectural thinking despite AI assistance.

6. Synthesis and Conclusion

The main takeaway is that Open Claw represents a shift toward agentic autonomy where users maintain control over their own data. While the project faces significant hurdles—specifically regarding security, the burden of maintaining a massive open-source codebase, and the "fear-mongering" surrounding AI agents—Steinberger remains committed to a decentralized, foundation-led model. The future of the project lies in making agents ubiquitous (e.g., "Star Trek-like" home integration) while teaching engineers to prioritize system design, "taste," and the discipline to say "no" to feature creep.