Peter Steinberger built the initial version of OpenClaw in just one hour, after wanting a seamless way to send prompts to his computer while away from his desk. "Anytime I get annoyed, I come up with a good idea," he said.

Originally named Var Relay, the open source project evolved rapidly by adding image support and dynamic capabilities. Within six months, OpenClaw became the fastest-growing open source project in software history, amassing 250,829 GitHub stars in just 60 days and surpassing the 10-year record previously held by React.

The project relies heavily on autonomous agents to learn, modify and orchestrate code dynamically with minimal human intervention. One surprising outcome of the release was the volume of pull requests from first-time contributors. "What surprised me the most was how many people sent the first PR ever on the OpenClaw repository," Steinberger recalled. Because the software was designed with self-awareness, early users often asked it to build its own missing features. The repository now has more than 2,000 unique contributors, many of whom are not traditional software engineers — something Steinberger sees not as a quality problem but as a signal of real user intent.

To keep quality high without constant manual review, Steinberger built an automated process he calls "order review," in which the underlying agent runs a command-line tool to review its own changes in a fresh session until no further edge cases are found.

Steinberger argues the evolution of coding agents reflects a broader shift across all digital work. Training agents to write code effectively also trained them to solve general problems, making the tools useful well beyond programmers. "Now pretty much any task is better with agents," he said.

Users interact with OpenClaw through familiar chat apps like WhatsApp, simply describing what they want built. "Everyone should use agents, not just developers," Steinberger said, describing how people could message a "friend" on WhatsApp and have software created for them without ever realising code was being written.

Verifying complex, agent-written software remains difficult for non-technical users — something Steinberger calls "the hardest part." His solution is automated, isolated testing environments across Linux, Windows and macOS, where an agent installs and runs the software it just built on a fresh machine. Combined with browser automation and computer vision, agents like OpenAI's Codex can now take screenshots of an app's interface to confirm that generated features actually work as intended.

Steinberger ranks the business impact of AI software agents above the mobile revolution, comparing its reach to the early internet, and notes that the millions of developers already using coding agents represent only a tiny fraction of the world's knowledge workers. "Even if we don't create better models in the future, there's so much that we can still improve and automate," he said, adding that "the main limitation right now is imagination."

To describe today's models, Steinberger uses the term "clankers" — a nod to the battle droids of Star Wars — cautioning against humanising agents that display what he calls "spiky intelligence": excelling at complex logic while failing at basic tasks. "They are an amazing tool everyone should use, but they are not people," he said.

Despite the ubiquity of automated coding tools, Steinberger insists human direction and domain expertise remain essential. "If you don't deeply understand software, you will not be able to build something very complex and have a good time maintaining it," he stated. He believes foundational computer science concepts, like algorithm efficiency, remain vital, and that without domain knowledge, users will fail to prompt agents correctly, leading to poorly performing software. As he put it: "Taste is the hardest thing that you can automate."

Key Takeaways

• →OpenClaw began as a one-hour side project and became the fastest-growing open source project in history, gaining 250,829 GitHub stars in 60 days and beating React's decade-old record.
• →Autonomous agents now learn, modify and orchestrate code with minimal human intervention, and Steinberger's "order review" process lets agents self-check their own work.
• →Agentic AI is expanding beyond coding to general problem-solving, letting non-developers build software through chat apps like WhatsApp.
• →Isolated testing environments and computer vision now let agents install, run and visually verify their own generated software across operating systems.
• →Steinberger calls today's models "clankers" — powerful but not human — showing spiky intelligence that excels at some tasks while failing at simple ones.
• →Despite rapid automation, human taste, domain expertise and architectural judgment remain irreplaceable in producing complex, maintainable software.