Steven Levy published a 5,000-word feature in WIRED on Monday morning that attempts something no outlet has managed yet: a single, authoritative account of how Claude Code and OpenClaw kicked off the current agent era. The piece opens at a London meetup called “Claude Code Anonymous” in August 2025 and follows two parallel origin stories to their convergence point.

Two Timelines, One Inflection

Levy traces Claude Code’s development to early 2024, when Boris Cherny left a remote engineering role at Instagram in rural Japan to join Anthropic. An engineer named Adam Wolff showed Cherny the company’s “very primitive” coding automation. Cherny used it to submit a pull request. “It wasn’t a good PR,” Wolff told Levy. But the attempt proved that better pull requests, and eventually higher-level coding tasks, were possible.

The team shipped a preview in February 2025, launched in May, and then released Opus 4.5 in November. That release introduced longer runtimes, better problem-solving, and subagent orchestration. Anthropic’s own internal benchmark claimed Opus 4.5 “scored higher than any human candidate ever” on the company’s engineering hiring exam.

The second timeline belongs to Peter Steinberger. After selling his company shares in 2021, Steinberger spent four years in what he described in a blog post as “hunting hedonic pleasures.” He discovered Claude Code’s beta in April 2025. By November, frustrated that the tool required terminal monitoring, he built a prototype agent accessible via phone, using OpenAI’s Codex. “I just prompted it into existence,” he told Levy. He released it as open source under the name Clawd (later renamed OpenClaw at Anthropic’s request) in late November 2025. It hit 100,000 GitHub stars in under two weeks. As of early May, it stood at 366,000.

The Productivity Claims

Levy’s reporting surfaces specific numbers from named executives. Garry Tan, Y Combinator’s CEO, told Levy he was coding at roughly 4 million lines per year with Claude Code, approximately 90x his best output as an engineer in 2013. Weeks after the interview, Tan revised that estimate upward to 408x.

Ryan Petersen, CEO of Flexport, described spending less time on C-suite responsibilities and more time building with Claude Code. Levy observed that the global supply-chain crisis in the Strait of Hormuz seemed to Petersen “less a corporate emergency and more an unwelcome distraction from his sessions with Claude.”

Cherny himself reported running “dozens, sometimes hundreds, of agents running eight and 12 hours at a time” most nights, with some agents operating continuously for days to rewrite codebases or optimize code efficiency.

The “Agent of Chaos” Paper

The piece doesn’t ignore the risk side. Levy cites a February paper by 20 AI researchers who tested OpenClaw and documented “unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions.” He also reports that a Meta safety and security engineer made a “rookie mistake” in an OpenClaw project and watched her inbox begin deleting all her mail.

This is the same research that Chuck Russell analyzed on Medium in March, noting that “OpenClaw agents take actions appropriate to Mirsky’s L4, while operating with L2 levels of understanding.” The gap between capability and comprehension is the core safety problem the paper identifies.

The Foundation Play

The piece also documents how Dave Morin, former Facebook executive and current VC, installed OpenClaw in December 2025 and immediately began a collaboration with Steinberger. “You’ve uncovered the Linux of AI, and it’s going to be 6-billion-people scale,” Morin told Steinberger in a January DM. Together they co-founded the OpenClaw Foundation, positioning the project as an open-source exemplar of beneficial AI.

What the Piece Actually Accomplishes

Most agent coverage to date has been either product reviews or policy commentary. Levy’s feature is the first major longform piece to treat agent adoption as a psychological phenomenon. The recurring framing is addiction: “Claudeholic,” “Claude Code Anonymous,” Steinberger’s insomnia, Petersen neglecting his supply chain, Cherny running agents through the night. Thomas Reardon, former Microsoft and Meta executive, tells Levy this is “the most underrated, massive release I’ve experienced in technology.” Marc Andreessen, on a recent podcast, said the quiet part: “It’s almost inevitable that this is the way people are going to use computers.”

The piece’s implicit argument is that agent adoption is being driven less by rational cost-benefit analysis and more by a compulsive, almost biological pull that technical users feel when they first experience autonomous task completion. That framing matters for anyone trying to predict adoption curves, because it suggests the bottleneck is exposure, not persuasion. Once someone uses an agent, the decision is already made.

That also explains the safety gap. The 20-researcher paper found serious vulnerabilities. Users who reported problems acknowledged them. And adoption kept accelerating anyway. The question Levy’s piece raises without answering: what happens when agents that were adopted compulsively reach the scale Morin envisions?

Published May 26, 2026, 12:15 UTC.