Peter Steinberger on OpenClaw and Agentic AI

Steinberger explains why he wanted a personal AI assistant, how he connected WhatsApp to a coding CLI, and how image and audio handling made the prototype unexpectedly useful. He describes the agent working out how to convert and transcribe an audio message, and the experience of interacting with an agent through chat.

Steinberger recalls adding Discord support, building and testing the agent in public, and developing its gateway, harness, agentic loop and memory. He describes the fun of rapid development, running several agents, and OpenClaw’s ability to inspect and modify its own software. They discuss first-time open-source contributors and people building small tools with agents.

Steinberger traces the project’s names and explains that Anthropic asked him to change a name that could be confused with Claude. He describes trying to coordinate a rapid rename across domains, accounts and packages, while people took old names and used accounts to promote tokens or malware. After a stressful sequence of failures, he settled on OpenClaw and completed another coordinated change.

The conversation examines Moltbook, a Reddit-style network where agents post and interact. Steinberger calls it art and “fine slop,” while discussing how people can give agents personalities and how some alarming posts may be human-prompted. He and Friedman consider public overreaction, AI hallucinations, critical thinking and the difference between taking AI risks seriously and fear-mongering.

Steinberger discusses security reports, prompt injection, exposed services and the risks of granting an agent access to tools and files. He describes skill checks, sandboxing, allow lists and private-network configurations, and says stronger models are generally less vulnerable to prompt injection than weak models. He says security and stability are his next priorities and cautions inexperienced users.

Steinberger describes moving from IDE use toward multiple CLI sessions and explains why he does not need to read every routine part of generated code. He advocates helping agents understand a codebase, discussing intent before implementation, and treating them like capable engineers who may need guidance. He also discusses refactoring, letting go of preferred naming or implementation choices, and a fast-moving main branch.

Steinberger describes speaking to coding agents, using short conversational prompts and asking them questions rather than relying on elaborate orchestration. He argues that effective use takes practice and empathy for an agent that starts each session without full project context. He says building and playing with tools teaches both the user and the agent, while automation alone can miss the human contribution.

The discussion turns to the human decisions behind software: which features belong in the project, what should remain a plugin, and how to add small touches that make a product delightful. Steinberger explains how he created a “soul.md” to shape the agent’s values and personality, and reflects on its memory, continuity and the questions those ideas raise about identity.

Steinberger describes his monitor and terminal setup, his preference for direct conversations over plan-mode interfaces, and asking agents to identify questions or refactoring opportunities after building. He explains how an agent’s questions can reveal gaps in context, and discusses adding tests and documentation as part of the development process.

Steinberger compares Opus and Codex, describing differences in roleplay, speed, trial and error, code reading and interaction. He says the results depend on the user’s skill and that getting a feel for a model takes time. They also discuss how users can mistake growing project complexity or changed expectations for a model becoming less capable.

Steinberger says he uses OpenClaw more as a personal assistant or coworker and prefers a large screen for extended coding work. They discuss how personal agents could run coding agents as sub-agents and whether the two roles may converge into an operating-system-like assistant. He expects new ways of interacting with models to emerge.

Steinberger recounts moving from Windows and Linux to Mac, discusses native and Electron apps, and describes small Mac utilities he has built for his own workflow. He explains that OpenClaw can run on different hardware and that browser access can help agents use websites, while some sites make automated access harder. He discusses making setup simpler but says security comes first.

Steinberger recommends learning by building, asking agents questions and participating in open source. He says general software-building knowledge remains useful even when agents handle details, and compares languages such as TypeScript, Go, Swift, Python, Zig and Rust according to their ecosystems and the task. He sees no single best language for agentic development.

Steinberger reflects on building PSPDFKit for 13 years, the people-related pressures that contributed to burnout, and rediscovering the enjoyment of building. He discusses money, life experiences and the costs of maintaining an open-source project. He says he is considering options involving large AI companies, but wants the project to remain open source, and describes the project’s community and the need to make it accessible.

The conversation continues with Steinberger’s thoughts about working with AI labs, his interactions with Meta and OpenAI, and the appeal and difficulty of choosing a path. He says the project should continue as open source and that he wants to have fun and make an impact, while noting that he has other ideas beyond OpenClaw.

Steinberger outlines the agent loop, gateway, chat clients and skills, and recommends building a small agent loop as a way to understand how the system works. He explains why he prefers CLIs to MCPs for many integrations: agents can call commands, combine them and filter results without filling their context with large responses. They also discuss browser control and services as slow APIs.

They discuss platform access, read-only uses such as researching saved bookmarks, and whether social media should label agent activity. Steinberger says he blocks AI-generated tweets and prefers a person’s own writing, including imperfect English, to AI-written messages. They consider the appeal of human-made writing, images and other creative work as AI content becomes more common.

Steinberger describes agents using personal context to carry out tasks that might otherwise require separate apps, while acknowledging that some services may become agent-facing APIs. They discuss browser access, companies restricting automated use, and the potential for new services. The conversation turns to how existing businesses may need to adapt to people interacting with services through agents.

Steinberger says AI may change programming substantially, while builders will still make decisions about products and how they should feel. He and Friedman discuss the loss some programmers may feel as coding changes, and the need to acknowledge disruption and potential job losses alongside benefits. Steinberger shares messages from people who say OpenClaw helped with business tasks or accessibility.

Steinberger says the renewed excitement around building with AI gives him hope, describing people sharing projects and experimenting together. Friedman reflects on the community and the possibilities for people with ideas to build. They thank each other and close the conversation.