One summer during high school, I took a temporary job writing computer programs for a consulting firm. Each morning, I drove through rush-hour traffic to an office park near Princeton, New Jersey, on the crowded Route 1 corridor. At a desk in some sort of equipment room, I coded quick-and-dirty database tools for internal use. One of my programs simplified the process of logging hours into timesheets. Another tracked inventory for the I.T. department. My role was to find small ways to improve the lives of my co-workers.

I hadn’t thought about that job in a long while, but it came to mind recently as I grappled with the potential economic impact of artificial intelligence. Proponents and critics of A.I. often compare the technology to industrial automation: just as machines eliminated many jobs that depended on human brawn, such as weaving or mining, A.I. will eliminate jobs that require human brains. In February, Mustafa Suleyman, the C.E.O. of Microsoft AI, predicted that A.I. will deliver “human-level performance on most if not all professional tasks” within the next twelve to eighteen months. Anthropic’s C.E.O., Dario Amodei, has suggested that fifty per cent of entry-level white-collar work will be automated by the end of the decade. “A lot of jobs will go away,” Sam Altman, the head of OpenAI, said last summer. (OpenAI has an agreement with Condé Nast, the owner of The New Yorker, which allows OpenAI to display its content in search results for a limited term.)

Recently, however, I’ve sensed a shift in this discourse. “A.I. is creating jobs,” the C.E.O. of Nvidia, Jensen Huang, said in an April interview. “Anybody who is saying that A.I. is wiping out jobs is scaring people.” He later said that C.E.O.s who have blamed layoffs on A.I. are lazy, and that such claims were “just a way for them to sound smart.” Even Altman said in May that he was “delighted to be wrong” about A.I. eliminating large numbers of jobs. Public-relations concerns are probably playing a key role here: eighty per cent of Americans are now somewhat or very concerned about A.I., according to a recent Quinnipiac poll, and a majority think that this technology will cause more harm than good in their daily lives. But, more fundamentally, we may be starting to realize that the automation analogy was never the right way to explain A.I.’s impacts.

In search of a better understanding of our current moment, I recently went looking for A.I. adopters outside the tech industry. I asked the C.E.O. of a journalism nonprofit how he’s using A.I., and he showed me a web-based tool that he vibe-coded using Claude Code, Anthropic’s programming agent. Each morning, the tool automatically summarizes articles related to higher education, suggesting potential trends and angles that could warrant further investigation. It then e-mails a brief to him and his managing editors. A recent brief highlighted a Los Angeles Times story about a data breach of a popular learning management system called Canvas. It suggested that the editors consider sending a Freedom of Information Act request to state school systems that were impacted, asking for correspondence with the system’s parent company. “Did anyone raise red flags?” it asked. The tool is hardly revelatory, and the C.E.O. said that he “would never try to turn it into a public-facing product,” but it highlights useful information and sparks ideas. “It’s like a student or an intern,” he said.

Lately, the C.E.O. has been thinking about another inefficiency that A.I. could address. His reporters fill out regular forms to summarize the impact of their work to send to the organization’s funders. Because he has access to Claude Code, he began imagining a bot that would solicit this information in a more informal way. Perhaps reporters could type their updates directly into Slack, the messaging platform that they already use, and the bot could fill out the form on their behalf. “It won’t be hard,” he told me. When I followed up with him a couple of weeks later, he confirmed that he had indeed created a tool to help reporters draft communications with funders. (It wasn’t yet integrated with Slack.)

Another A.I. enthusiast, the co-owner and chairman of a shipping-logistics company, told me about a “big headache” that was vexing his C.F.O. “We get payments from thousands of clients,” he said. “Often, they don’t note what they’re paying us for.” The company had given up on changing clients’ behavior; instead, four staffers were assigned to match mystery checks to corresponding invoices. (I couldn’t help but think of a Sisyphean number-filing task in “Severance,” the dystopian Apple TV+ series set in a workplace.) But, earlier this year, the company gave the I.T. team access to A.I.-powered coding agents. Staffers quickly built a custom tool that automated “essentially eighty per cent” of the matching issues on a recent project, the co-owner said. He was now in the process of reassigning three-quarters of his human payment-matchers to more fruitful tasks. These examples were not the digital equivalent of a power loom, making large numbers of human jobs superfluous. Turns out, A.I. was assisting these small businesses in roughly the same way that my teen-age self had helped that consulting company—by hacking together whatever was useful.

The ability to vibe-code custom software using A.I. might be new, but it actually echoes a much older vision for personal computers. In the seventies, when these machines were first introduced, easy-to-learn programming languages such as BASIC—meaning “beginner’s all-purpose symbolic instruction code”—were meant to empower any user to write their own programs. A nineteen-year-old Bill Gates, alongside his friend Paul Allen, developed a version of BASIC to run on the Altair 8800, the very first commercially viable personal computer; they soon formed a company together that they called Micro-Soft. Steve Wozniak created his own version of BASIC for the Apple I, which he and Steve Jobs released a year later. The Apple II shipped with a copy of BASIC hard-coded into its memory chips. “That means you can begin writing your own programs the first evening, even if you’ve had no previous computer experience,” an ad declared.

This idea of bespoke computer programs made sense. Altair and Apple couldn’t anticipate every potential use for their machines, so why not let individuals decide whether they wanted to, say, analyze business data, store recipes, or simulate space battles? In practice, however, even an “easy” programming language like BASIC proved hard for most normal people to master. A minor mistake could crash an entire program. In the end, personal computing followed a different path. In 1979, a newly formed company called Software Arts developed VisiCalc, the first electronic spreadsheet program, which cost a hundred dollars and arrived on a floppy disk. The program was a profound improvement on paper ledgers, and it became the first “killer app,” selling more than seven hundred thousand copies in less than six years. VisiCalc was more powerful than anything an average user could program in BASIC, and it prompted a pivot away from D.I.Y. coding in favor of professional programs. A vast and lucrative software industry emerged, and the idea of the average person dreaming up their own custom programs was all but forgotten—that is, until generative A.I. came along. Arguably, the nonprofit C.E.O. and the shipping executive were returning to the original vision of custom computation.

I pitched my theory—that A.I. is not so much replacing workers as equipping them with bespoke tools—to Taylor Pearson, a consultant who writes a popular newsletter called “The Interesting Times,” about the intersection of business and technology. He shared my skepticism about the idea that artificial intelligence is causing some sort of workplace apocalypse. At a recent conference for small-business owners, for example, he didn’t hear much about A.I. taking jobs; the next closest thing, he said, was that some companies were cancelling I.T. contracts because chatbots were helping them solve simple tech problems. But he agreed that vibe-coded custom software is increasingly ubiquitous. He gave the example of a company that rents dumpsters. It recently replaced a commercial platform that managed its customers with a vibe-coded version that was better at detecting fraud. (Apparently, it’s common for fraudsters to offer dumpsters at a big discount, collect a client’s money, and then rent an actual dumpster using a stolen credit card.) But Pearson thought that my theory didn’t go far enough. In recent months, he has been writing about a new style of A.I. collaboration. He doesn’t think it’s going to eliminate our jobs—but, as soon became clear when we spoke, it certainly has the power to transform them.

Earlier this year, Pearson needed to replace his house’s central air-conditioning system, so he obtained quotes from several HVAC companies. He told me that, to him, the proposals were “all gobbledygook.” He decided to fire up Claude Code—not to write software but to solve the real-world problem of deciding which A.C. unit to buy. A.I. agents like Claude Code differ from standard chatbots in a couple of main ways. First, they combine a standard large language model with a sophisticated control program, sometimes called the harness, that can execute multiple steps on the user’s behalf. (The harness might ask the L.L.M. to create a step-by-step plan, request its help in executing each step, and then check the validity of the results.) The harness may also have permission to read and write files, and to access certain tools, on the user’s computer. “It gives the model ‘hands’ to be able to do things,” Pearson told me. In the process, A.I. becomes less like an oracle that answers questions and more like a work partner.

Pearson gave the coding agent access to a folder containing the PDFs of all the quotes he had received. He asked it to read them and identify the specific models of A.C. systems that they referenced. The agent then searched the internet, found each system’s user manual, and drafted a report on their features. As the agent worked, it created a useful knowledge base in the form of subfolders and text files on Pearson’s computer. When he asked if a particular unit would fit between his house and his neighbor’s fence, for example, the agent had easy access to the relevant information.

Pearson didn’t trust Claude Code to make the decision for him, though. These agents “do and say a lot of boneheaded things, and change their minds a lot,” he said. This means that they require close supervision and are best at more mechanistic tasks. Pearson estimates that it took twenty different chat sessions to complete the A.C. project. But his decision was far better informed than if he’d simply picked one of the quotes—and easier to make than if he had attempted to read all of those A.C. manuals himself. He now uses Claude Code to support everyday tasks on a regular basis. Recently, it helped him research and compare fifty different exchange-traded funds—a complicated endeavor that drew on data from multiple websites and involved using a script, a simple computer program, to calculate some statistics. He also uses Claude Code to help edit his newsletter, sometimes by giving it verbal commands such as “I want to move this piece up and this piece back, and take this section out and use it in this next piece.” The A.I. isn’t doing Pearson’s job for him, but it helps him do the job better. “The percentage of my time spent on more cognitively demanding and interesting things is probably up somewhat,” he said. Pearson calls this approach “freestyle work” because it reminds him of a type of freestyle chess, in which humans collaborate with computers to play the game at a higher level.

It’s not yet clear that freestyle work will come to your workplace anytime soon. (Anthropic seems to think it will: earlier this year, it released Claude Cowork, a more office-worker-friendly version of Claude Code.) But vibe-coded custom software is so useful, so tailored to our specific needs, that its spread is almost inevitable—a reality that seems to be reflected in sharp stock declines experienced by web-based software companies. A.I. really is changing work—and if we all start barking commands at our personal A.I. agents, our jobs could get a lot weirder. But, so far, we don’t seem to be hurtling toward the darker vision of a workforce hollowed out by this technology. My summer job in New Jersey, all those years ago, didn’t put any consultants out of work. I like to think that it just made their efforts a little deeper. ♦