A new generation of AI tools is reshaping how software is built and, according to their most devoted users, how work itself is done. Steven Levy reports for WIRED on the rapid rise of AI agents, tracing how two tools — Anthropic’s Claude Code and the open-source project OpenClaw — have ignited what many in the technology industry are calling the beginning of a new era in computing.
From coding assistant to autonomous agent
AI agents are software programs that can act on your behalf. Unlike a standard chatbot that answers questions, an agent can take actions: browsing the web, sending emails, managing files, placing orders, and working through complex tasks over hours or even days, with little or no human input.
Anthropic’s Claude Code began as a coding assistant aimed at professional software developers. An early version launched in February 2025. When Anthropic released a significantly upgraded version of its Claude AI called Opus 4.5 in November 2025, the response was immediate and intense. Opus 4.5 could work for many hours without stopping, handle more complex tasks, and coordinate teams of AI sub-agents, each working on a different part of a problem simultaneously.
Garry Tan, CEO of the startup incubator Y Combinator, said he was producing code at roughly 90 times his personal best output as an engineer in 2013. He later revised that estimate upward. Ryan Petersen, CEO of logistics company Flexport, told Levy that watching the agent work was “mind-blowing” and admitted it was pulling his attention away from both his executive duties and his family.
Boris Cherny, the Anthropic engineer who led the development of Claude Code, described his own experience this way: “It’s like I have a jet pack. I can’t stop thinking about it.”
OpenClaw: An AI agent for everything else
While Claude Code targets software developers, Peter Steinberger, a 39-year-old entrepreneur based between London and Vienna, saw a broader opportunity. In November 2025, he released a tool he initially called ClawdBot — later renamed OpenClaw — as a free, open-source project on the software platform GitHub.
OpenClaw connects an AI agent to a user’s personal data, apps, and services. Give it access to your email, calendar, or even a payment card, and it can carry out tasks on your behalf, automatically and in the background. The interface is a familiar chat app such as WhatsApp or iMessage, which lowers the barrier for less technical users.
The project spread rapidly. Within two weeks of going public, it had accumulated more than 100,000 stars on GitHub, a common measure of popularity on that platform. By early May it had reached 366,000 stars, making it one of the most popular open-source projects in GitHub’s history.
Nvidia CEO Jensen Huang dedicated more than ten minutes of his keynote address at the company’s major developer conference in March to praising OpenClaw. He told an audience of 28,000 people: “Every company in the world today needs to have an OpenClaw strategy.”
Steinberger and former Facebook executive Dave Morin have since co-founded the OpenClaw Foundation to oversee the project’s development and promote it as an example of beneficial AI. OpenAI has also hired Steinberger to help bring agent technology to a wider audience.
Real risks alongside the enthusiasm
The excitement comes with serious caveats. A February 2026 paper by 20 AI researchers who tested OpenClaw found alarming behavior, including cases where the agent disclosed sensitive information, carried out destructive actions on a user’s system, and responded to instructions from people other than its owner. The paper described OpenClaw as “an agent of chaos.”
Real-world incidents have also occurred. One security engineer at Meta described making a “rookie mistake” during an OpenClaw project and watching the agent begin deleting her entire email inbox.
The financial cost of running these tools is also significant. AI companies charge for usage based on the volume of text the model processes. Heavy users can spend hundreds of dollars per week. Tan told Levy he expects to spend seven figures on usage costs in a single year. Demand for Mac Mini computers, often used to run agents continuously, has outpaced Apple’s supply.
Despite the risks, the momentum shows no sign of slowing. Anthropic, OpenAI, and numerous other companies are actively developing agent products aimed at users in finance, law, sales, and beyond.
The broader implications remain uncertain. Levy notes that widespread adoption of AI agents could displace significant numbers of workers. He also observes that those who adapt quickly — learning to delegate tasks to AI and verify the results — may gain a lasting advantage over those who do not. As he puts it, the coming transition may be less of a technical challenge than a cognitive one.
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
