OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger

OpenClaw's first useful form joined WhatsApp to a coding-agent CLI: receive a message, invoke the tool, return the answer. Steinberger says the basic relay took an hour, with image support requiring additional work. Earlier experiments included querying WhatsApp history and exposing terminals on the web. The apparent overnight breakthrough therefore reused accumulated tools and learning rather than beginning from nothing.

A trip to Marrakesh made the interface's value concrete. WhatsApp worked on weak connections, while the agent could translate, explain and find places using existing command-line tools. Fridman distinguishes messaging an assistant from sitting at an IDE: the small interface change alters how AI fits into daily life. Their comparison with iPhone scrolling concerns the experience created by combining familiar components, not a claim to have invented every ingredient.

Fridman frames the project as a move from language to agency, with access to personal data supplying both utility and risk. The biographical parallel is PSPDFKit, which began with frustration at displaying a PDF on an iPad. Steinberger's recurring response to an inadequate tool is to build one, while the conversation's opening jokes distinguish deliberate agentic engineering from late-night experimentation that needs repair the following day.

An unplanned voice message became a turning point. Steinberger recounts the agent identifying an extensionless Opus file, converting it with ffmpeg and using an available OpenAI credential when local Whisper was unavailable. He interprets this as coding ability transferring into general problem solving. It is a reported incident, also revealing how broad access enables actions beyond an explicitly implemented feature.

Discord support made public demonstration easier than sharing a phone number. The early bot had no sandbox and relied on a prompt to obey its owner; that history is not a security endorsement. Development expanded through message queuing, an option not to reply in group chats and file-based memory. Steinberger describes the work as an expanding playground, with several coding agents running in parallel, rather than a fully planned architecture delivered at once.