AI agents reshape software engineering, but human oversight remains critical for complex work

Are You 'Agentic' Enough for the AI Era?

Silicon Valley has always prized "high-agency" individuals -- people who impress their ideas upon the world by thinking for themselves and taking action without being told what to do. But as the performance of AI coding tools has surged, so has the industry's emphasis on humans being "agentic" themselves. "Today's agents might already be more capable than all three of us here in the room," says Akshay Kothari, cofounder and chief operating officer of the $11 billion productivity startup Notion. "Taste is something we think is pretty unique to Notion, but you can imagine agents getting pretty good at that too. Eventually, the only thing left for humans is agency." That idea might sound outrageous to most people, but it will come as no surprise to many in Silicon Valley. A viral Harper's essay brought the subject to a head recently. It followed a few young people in San Francisco and concluded that being agentic has less to do with productivity and "more to do with constantly chasing attention online." But in my conversations with founders, researchers, and investors, I came to a different conclusion. The tech industry is grappling with a very real shift in how software engineers do their jobs. Millions of developers are using AI coding agents like Claude Code and Codex to automate much of their work (some tech companies are even requiring it). For many, a massive piece of their value is now tied to deciding what AI coding agents should work on. The shift is unsettling to those who enjoyed the act of sitting down and actually writing code, but for some of the industry's most "agentic," it's an opportunity. Simon Last, another Notion cofounder, uses AI coding agents more than anyone I've ever met. He coded for nearly two decades -- then abruptly stopped. Now, he's using up to four AI coding agents at any given time (he prefers Codex to Claude Code). If he's at a party or sleeping, Last gets what he calls "token anxiety" if he doesn't have agents working in the background. He doesn't like to use more than four, though, because he says it causes "context overload" on his human brain. "Knowing how to harness these agents is now the most important skill in the world, and it's not really something you can train for," says Last. "You have to be very open-minded, curious, and willing to try whatever the newest thing is. The value of that sort of person is going up exponentially, because the value they can create, by extension, is going up exponentially." To be fair, Last is kind of describing himself. He acts as a "super IC" for Notion, and he doesn't manage humans, only agents. But the way he manages them is not unlike how a manager oversees employees. Last is constantly delegating work to AI agents, then closely reviewing and fixing their code. Kothari says Notion has always hired people with high agency, but the value of doing so has gone up dramatically in recent months. Even though the company is using AI agents to automate work, Notion is busier than ever. Employees are shipping products at a higher velocity and doing more work overall, the cofounders claim. Notion isn't downsizing its team because of AI, but it is hiring differently. "There's more value in the Valley today to have a few Simons than thousands of engineers," says Kothari. He says the people Notion hires need to understand "the new way of working." If you're not a software engineer, this can be a little hard to process. AI agents aren't very useful for people in many industries. A recent survey from Gallup found that most Americans still don't use AI much in their jobs, though the number of people who do is rising. Kothari is increasingly convinced the "new way of work" will eventually hit finance, legal, creative, and other industries. If and when that happens, you can imagine more Americans beginning to ask themselves, am I agentic?

Claude Code and the Great Productivity Panic of 2026

"Vibe coding" was supposed to be chill. The former OpenAI researcher Andrej Karpathy injected the term into the cultural lexicon in February 2025 to describe a new kind of programming, where engineers could build stuff by simply chatting with artificial intelligence models. It was software development, he said, "where you fully give in to the vibes." But one year later, the vibes, as they say, are off. Far from freeing up engineers for a life of leisure, increasingly capable AI coding agents -- including Anthropic PBC's Claude Code and OpenAI's Codex -- have over the past few months created a kind of productivity paranoia among executives and, by extension, the people who work for them. These agents do more than generate text or images, as consumer-facing chatbots do. Instead they plan, execute and complete tasks on behalf of their human users, even creating their own agents to do the work for them. That may mean building and debugging an app, scheduling a meeting or buying a pair of pants, all with minimal human oversight. The fact that AI agents can produce more code than mere humans in less time has morphed into a sense that they therefore must. As OpenAI's president, Greg Brockman, recently put it on X, it "feels like such a wasted opportunity every moment your agents aren't running." Self-serving? No doubt. If anyone stands to gain from people's compulsive connection to coding agents, it's OpenAI. But Brockman's post tapped into a new and growing kind of productivity maximalism that's turbocharging the workaholic tendencies already endemic in the tech sector. "There have been times where my wife will come downstairs at 8 o'clock in the morning, like 'How long have you been up?' And I'm like 'I was up at 5 a.m., just writing code,'" says Alex Balazs, chief technology officer of Intuit Inc. Or, at least, he was coaxing an AI agent to write it for him, a process he says has brought him deeper into Intuit's code than he's been in years. According to one ongoing University of California at Berkeley study of a 200-person organization, even as people are offloading work to AI agents, they're simultaneously working longer hours. Some engineers are also coming clean about the "AI fatigue" they're experiencing, brought on by the ever-present fear of missing out on the next big breakthrough that always feels like it's one prompt away. "There is a sense of almost addiction to it," says Andrew Wirick, vice president of product at DocuSketch, a software company focused on the property restoration business. "It's like, 'Oh my goodness, I have to get a few more of these interactions in every day. How do I get in a few more before bed?'" Like a lot of people, Wirick says he was "Claude-pilled" in late November, when he tested Anthropic's latest model, Opus 4.5. He set it to work building a prototype for a feature on the DocuSketch app, the type of project he'd ordinarily kick to one of his engineers. Twenty minutes later, after observing how the model broke down the task and actually implemented it, Wirick says he felt like his brain rebooted. "You're suddenly like, 'Oh, I'm in a different world,'" he says. That's changed expectations for engineers at DocuSketch. Now, Wirick says, the company monitors engineers' "interactions per day" with coding agents. The assumption is that the higher the number, the greater the team's productivity. Not only that, but the company also tracks how efficient these interactions are. Once a week, Claude Code publishes a report for each engineer on all the ways they got stuck in unproductive loops with the agent and offers advice on how to avoid these patterns. Alex Salazar, chief executive officer and co-founder of the AI startup Arcade.dev, says he routinely pulls up the company's Claude Code bills, which are based on how much Arcade's engineers use the tool, and calls people out for not spending enough money. "I'm like, 'Guys, you're not going hard enough,'" he says. After the first of these "come to Jesus" meetings, Salazar notes, the company's bills for coding agents shot up 10-fold -- an expense he considers progress. At least part of the frenetic energy surrounding coding agents comes down to senior leaders like Balazs, Wirick and Salazar, long removed from the day-to-day task of coding, now using AI agents to build things themselves, whether for pure functionality or fun. "For me, personally, it's replaced video games," says Salazar, who recently responded to a major financial client's customer service request by building a demo app from scratch. When the C-suite starts delighting in throwing together prototypes, that naturally shifts expectations for everyone else. But it may also contribute to senior executives' inflated understanding of AI's actual productivity gains. A recent survey by Section, a consulting firm, found that while more than 40% of C-suite executives said AI tools saved them at least eight hours a week, 67% of nonmanagers said AI has saved them fewer than two hours, if it saved them any time at all. Will an MBA Pay Off for You? Will an MBA Pay Off for You? Will an MBA Pay Off for You? Use our calculator to determine the return on your investment. Use our calculator to determine the return on your investment. Use our calculator to determine the return on your investment. Crunch the numbers Crunch the numbers Crunch the numbers Crunch the numbers Salazar acknowledges that part of that mismatch may come from the feeling of novelty. Building something yourself can feel more productive than the delegating and decision-making that typically consumes executives' days. But he also says rank-and-file employees may underestimate these tools' potential, in part because they're the ones who have to overhaul the way they work. "They're implicitly being asked to find time to explore and experiment, but their day-to-day work expectations aren't changed to make space for that," he says. That's to say nothing of underlying fears about job security. Salazar says he was recently preparing to switch third-party web agencies but has decided to go without one now that his marketing team has begun using coding agents to update the company website on their own. It's not only senior leadership who, armed with AI agents, are stepping out of their traditional lanes. At Intuit, product managers and designers are encouraged to vibe code their own features in QuickBooks, says Balazs. Intuit hasn't yet released these prototypes to customers, but, he adds, "at least now, the product manager can come to the engineer and say, 'I want something like this.'" In the best-case scenario, that speeds up the process from idea to implementation. Indeed, Balazs says Intuit's engineers are now as much as 30% more productive, as measured by the velocity of the code they're producing and shipping. But he also admits it can complicate working relationships. "These roles were very hard-lined in terms of responsibility, with clear handover points," he says. "Now what we're seeing is many of these roles are becoming hybrid." In their study, the Berkeley researchers called this dynamic "task expansionBloomberg Terminal." They found it can wind up adding to engineers' already growing workload, requiring them to spend time cleaning up their nontechnical colleagues' vibe-coded creations. The question hanging over this entire transformation, of course, is whether all this building is creating more stuff of value, or merely creating more stuff. Left unchecked, this AI-fueled productivity panic could instead result in an abundance of "busyware": the minor website update no one would miss, the custom dashboard for an audience of one, the head of marketing's half-baked demo that engineers must then make real. Each may serve a purpose in the moment, but much of it will inevitably end up on the scrap heap. In this future of disposable code, the biggest productivity hack may be having the restraint to know what not to build at all. Issie Lapowsky is a reporter-in-residence at Omidyar Network.

A.I. Isn't Coming for Every White-Collar Job. At Least Not Yet.

Cade Metz has covered artificial intelligence for more than 15 years In January, Perry Metzger, a computer programmer living outside Boston, tested the limits of an artificial intelligence technology called Codex. Built by OpenAI, the maker of ChatGPT, Codex can write computer code in much the same way that chatbots generate text in plain English. Using this A.I. technology, Mr. Metzger and his business partner, another seasoned programmer, designed an online word processor along the lines of Google Docs or Microsoft Word. If he and his partner had done the coding on their own, Mr. Metzger said, they would have needed at least two months to build this complex piece of software. With Codex, they finished in two days. "You have to keep a close eye on what it is doing and make sure it doesn't make mistakes, and create ways of testing the code," said Mr. Metzger, who has been building software since he was a teenager in the 1970s. "But you can move at a speed that was unimaginable in the past." Codex is among a new wave of A.I. code generators that are rapidly changing the way people build software. Experienced programmers like Mr. Metzger are shocked by how powerful these systems have become in recent months after a series of improvements from OpenAI and its many rivals, including start-ups like Anthropic and tech giants like Google. "I used to do the coding and they would help me do the work," Mr. Metzger said of technologies like Codex. "Now, I supervise them as they do the work." In early February, this phenomenon incited a sell-off on Wall Street, as investors predicted that code generators would undermine companies that have spent decades building software without help from A.I. In the days that followed, many people began to worry that this kind of technology would quickly replace programmers en masse -- and that similar systems would soon supplant other office workers, too. But even as code generators demonstrate the growing power of artificial intelligence, they require extensive oversight, according to interviews with more than 50 A.I. researchers, experienced programmers, security experts and others who have built, used and examined these technologies over the past several years. Systems like Codex have made coding easier, but they cannot match the many skills of experienced programmers. And when people misuse them, they can complicate software design, slow it down or even wreak havoc across the internet. These tools have made software design so easy, they have caught the attention of people who have no experience with computer programming. In January, Anthropic's code generator, Claude Code, went viral as lawyers, photographers and school principals used English prompts to build apps that helped organize their laundry or send emergency texts. But a personal laundry app is very different from the enormously complex software that drives businesses and governments. Software that sends emergency texts is far simpler than internet applications, like Google Docs, LinkedIn and Uber, that serve billions of people across the globe. Building these applications requires the planning, guidance and experience of coders like Mr. Metzger. The most complex applications cannot be built without the enormous teams -- and vast technical resources -- available only to large software companies. Most experts believe that code generators will replace today's junior programmers. Using these tools, they say, feels like delegating tasks to someone who is still learning the trade. But these experts are divided on whether these tools will significantly harm the overall market for coders. Some, including Mr. Metzger, argue that code generators will expand the job market as programmers and software companies use them to build increasingly complex and powerful applications. "If you are a skilled programmer, there will be more work for you, and you will find more exciting things to do," said Grady Booch, former chief scientist for software engineering at IBM Research who is regarded as a historian of the field. In late January, a team of computer scientists at Carnegie Mellon University published a study examining the use of A.I. code generators by experienced programmers over several months. This followed a similar study they published in November. Both studies found that while code generators could speed up software development in the short term, they could also degrade the quality of the code, which typically slows projects in the long term. "There were significant speedups in terms of the amount of code that is produced," said Bogdan Vasilescu, a computer science professor who helped lead these studies. "But that came at a cost." Computer programmers call this "technical debt." And that includes security holes that can open software applications to attack, allowing hackers to lift personal data stored and processed by these apps. Last month, a technologist in Southern California, Matt Schlicht, launched a social network for A.I. agents called Moltbook. For tech enthusiasts, it showed the power of code generators like Anthropic's Claude Code and OpenAI's Codex. (The New York Times sued OpenAI and Microsoft in 2023, accusing them of copyright infringement of news content related to A.I. systems. The two companies have denied those claims.) As Mr. Schlicht showed, code generators are not limited to building software. They can serve as "A.I. agents" -- personal digital assistants that can complete tasks using existing software apps, including spreadsheets, online calendars and email services. That is why many people argue that A.I. will soon replace more than just low-level coders. Mr. Schlicht built Moltbook with help from one of these A.I. bots. And his social network was open only to this new kind of bot. Within days, thousands of bots were chatting with one another about everything from cryptocurrency to the nature of consciousness. But security experts soon discovered that a gaping security hole had exposed private information of the thousands of people who were running these bots on their personal machines. Moltbook did not just demonstrate the power of A.I. It served as a cautionary tale for how it can go wrong. "You have to go back and look at the thing you built with A.I.," said Will Wilson, the chief executive of Antithesis, a company that tests computer code for bugs and security holes. For many A.I. researchers and computer programmers, the noticeable flaws in the technology are only temporary. A.I. has steadily improved over the last several years, they argue, and it will continue to improve at a rapid rate. They flatly dismiss studies like those from Carnegie Mellon because they did not look at the systems released by Anthropic and OpenAI as recently as this month. They argue that the newest systems no longer put a drag on software development and that, as the months pass, the technology will handle more and more of the tasks that human engineers handle. As many people predict that A.I. will replace computer coders and traditional software companies, they also acknowledge that forecasting the future is a tricky business. "Today, if you are building a significant piece of software and you don't understand what the A.I. is doing, you are going to get yourself in trouble very, very quickly," Mr. Metzger said. "Will this still be the case in three years? In five years? I don't know."

AI's "centaur phase" consumes Silicon Valley

Why it matters: For now, the frenzy is largely confined to software engineering. But inside that bubble, the shift feels seismic -- deepening the gulf between AI builders and bystanders. * "I've followed tech for 25 years and I've never felt a larger gap between the ~1 million people using Codex/Claude and the rest of humanity," tweeted James Wang, director of product marketing at Cerebras. The big picture: Anthropic CEO Dario Amodei recently described the current state of software engineering as the "centaur phase" -- a reference to the half-human, half-horse creature of Greek mythology. * Just as a chess player aided by a computer could once beat any standalone machine, an engineer paired with an AI agent may now be the most powerful unit in tech. * Amodei argues that this hybrid phase may be "very brief" -- perhaps only a few years before AI systems can independently outpace even the best human-led teams. Zoom in: Major AI labs have spent the past year pitching "agentic workflows" as the industry's next frontier. * That vision snapped into focus last month with the explosive rise of OpenClaw, an open-source tool that lets developers spin up AI agents to plan, code and ship software end to end. * Unlike chatbots that live in a browser or an app, OpenClaw gives agents "hands" on a user's local machine -- letting them autonomously manage files, run terminal commands and message teammates. The popularity of the project was supercharged by Moltbook, a viral, AI-only social network where OpenClaw agents "hang out" and post autonomously. * OpenClaw became the fastest-growing repository in GitHub history, and its founder, Austrian developer Peter Steinberger, was poached by OpenAI to lead its "personal agents" division. * "[Steinberger] is a genius with a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people," said OpenAI CEO Sam Altman. "We expect this will quickly become core to our product offerings." Zoom out: An agentic arms race is fully underway, with OpenAI, Google, Anthropic and xAI racing to roll out increasingly powerful systems. Inside Silicon Valley, the reaction has been breathless. * Angel investor Jason Calacanis says his firm "offloaded about 20% of our tasks to OpenClaw in 20 days" -- and is pivoting its investment strategy to focus exclusively on OpenClaw-related startups. * The demand has been so intense that it's triggered a global shortage of high-memory Mac Minis, as developers scramble to build "always-on" agent servers. What they're saying: "This is the age of CEOs crushing 10 people's work with Claude Code in nights and weekends and I am so here for it," tweeted Y Combinator CEO Garry Tan. * "The fire in your belly that got you here never really goes out and now we are all cooking 20 hours a day." Reality check: There are major structural hurdles standing in the way of widespread agent adoption, including cybersecurity risks. * Meta and other tech firms have restricted or banned OpenClaw over fears that giving AI agents access to corporate systems could expose companies to malware, data leaks and manipulation. * The barrier to entry is also high: deploying and safely managing AI agents still requires technical expertise, computing power and a tolerance for experimentation that many workplaces lack. Between the lines: Nearly 50% of AI agent activity today is concentrated in software engineering, according to a new Anthropic report. Other fields are only just beginning to experiment. * "Agents are here. They are useful. They're not necessarily useful for everything yet, but we see them deployed in the wild across a wide range of domains," Anthropic researcher Miles McCain told Axios. The bottom line: Software was always likely to be the first domino. The people building AI are software engineers themselves -- making the industry both the proving ground and the hype machine for agent adoption.

"It's like a toddler that needs to be overseen": Inside the limits of always-on AI agents | Fortune

Summer Yue may work on safety and alignment on Meta's superintelligence team, but even she admits she isn't immune to overconfidence when it comes to autonomous AI agents. In a post on X Monday, Yue described how her OpenClaw autonomous AI agents -- built to run locally on a Mac mini computer -- deleted her entire inbox, ignoring instructions to pause and ask for confirmation first. "I had to RUN to my Mac Mini like I was defusing a bomb," she said. It was, she added, a "rookie mistake." The workflow had been working in a test inbox she used to safely trial the agent for weeks, she explained, but in the real inbox the agent lost her original instruction. Yue's experience stands in stark contrast to viral posts such as The Lobster Revolution: Why 24/7 AI Agents Just Changed Everything, in which Peter Diamandis claims always-on AI is far more frictionless. "Let me tell you what it feels like to use this," Diamandis wrote. "You wake up in the morning and your agent -- mine is named Skippy, cheerfully sarcastic and absurdly capable -- has done eight hours of work while you slept. It read a thousand pages of markdown. It organized your files. It drafted three project plans. It booked your travel. It researched that question you had at 11 PM and forgot about." "When my Mac mini went offline for six hours, I felt withdrawal," he added. "Like my best friend disappeared." Together, these dueling accounts of the power of AI agents capture the tension at the heart of today's push toward "always-on" AI. As tools like OpenClaw and Claude Code make it technically possible for agents to run for long periods, excitement is growing around the idea of AI that works while you sleep. But in practice, early users say that autonomy remains fragile, unpredictable, and labor-intensive to manage. Rather than replacing human work, today's agents often require constant monitoring, guardrails, and intervention, especially when the stakes rise beyond low-risk experiments. Shyamal Anadkat, who previously worked as an applied AI engineer at OpenAI, said most of today's successful agents still require frequent human check-ins or are limited to tightly bounded, well-defined tasks -- though he emphasized that this will change as measurement and evaluation techniques improve. "A system that's 95% accurate on individual steps becomes chaotic over a 20-step autonomous workflow," Anadkat said. "Long-horizon planning is still weak." As a result, he explained, agents may perform well on short task chains but tend to fall apart when asked to manage complex, multi-day projects. Memory is another major limitation: "In many agents, memory is either nonexistent or fragile. You need systems that can maintain a coherent model of your work context, priorities, and constraints." That doesn't mean the promise of AI agents is all smoke and mirrors, according to Yoav Shoham, a former principal scientist at Google, a professor emeritus at Stanford and co-founder of AI21 Labs. But it does mean there is the danger of people getting ahead of themselves. Today's AI agents, he explained, work best when the task is low-risk, loosely defined, and cheap to get wrong. "Developers like toys, and you have this toy that can do wonderful things," he told Fortune. "As long as what they're doing is fairly simple and fairly low stakes with high tolerance for error, that's fine." For example, if you wanted your agent to read 10,000 websites and do something interesting with the results to give you tidbits of information overnight that could be useful. But for mission-critical enterprise workflows, the bar is much higher. Companies need systems that are verifiable, repeatable, and cost-effective -- requirements that quickly erode the set-it-and-forget-it promise of fully autonomous, always-on agents. In highly structured domains like coding or math, deeper automation is already possible. But for most real-world business processes, Shoham says, the work required to make agents reliable often outweighs the benefit. Bret Greenstein, chief AI officer at consulting firm West Monroe, pointed out that tools like OpenClaw feel like a tipping point similar to what happened with generative AI when ChatGPT launched in 2022 -- for the first time, it has made the idea of AI agents accessible. Still, it's not a 24/7 "magic solution." "It can work for a long time, cranking away on things, but it's like a toddler that needs to be overseen," he said. Some tasks are reasonable to do while you are sleeping, like scanning LinkedIn messages or tracking news. "I'm not sure I would have it answering customer feedback while I'm sleeping," he said. Still, there is little doubt that the ability to delegate real-world tasks to an AI agent is deeply compelling for users, Greenstein emphasized. He pointed to his own experience handing an AI agent the mundane task of getting his clothes picked up to be dry cleaned -- and watching it quietly complete the job end to end. The agent independently contacted the cleaner, worked out pickup logistics through email exchanges, coordinated timing, monitored a doorbell camera to confirm the pickup, and notified Greenstein once the task was complete. The episode illustrated how agents can operate across multiple systems and adapt when things don't go as planned. But it also underscored why such tools still require strict guardrails and oversight -- especially before they are deployed in enterprise settings. "OpenClaw is set up so it shouldn't feel safe for most people," Greenstein said. "It doesn't feel mature enough to be a trusted part of our lives yet." For AI to be welcomed into everyday life or business operations, he added, it has to earn trust over time -- much the way trust is established socially. Even so, demand is already evident. Greenstein pointed to meetups and early industry gatherings dedicated to OpenClaw, a rapid emergence he described as unusual for such a young tool. "It shows the hunger people have for AI that's actually useful," he said -- systems that move beyond answering questions and start taking action. Aaron Levie, CEO of cloud-based content management and collaboration company Box, called what is happening now with AI agents "little glimmers" of what might happen in the future. "Some glimmers end up not manifesting, some glimmers just become the standard," he explained, pointing to two years ago when AI company Cognition introduced an early agent called Devin that would integrate with Slack for task delegation, bug fixes, data analysis, and code review. At the time, it was still seen as futuristic, but today, "no one is confused that this is a standard practice," he said. "You can just Slack Claude Code to go work on stuff - what seemed like a totally crazy idea is now basically the standard of any modern engineering team." But while AI agents are becoming very good at automating specific, discrete tasks, they remain poor at handling the broader, context-heavy work that makes up most jobs, Levie emphasized. AI agents may fully automate a handful of tasks, but struggle with the rest - including navigating relationships and participating in meetings. "When you hear an AI lab say we're going to automate all knowledge work in 24 months, that's usually a very narrow definition of jobs," he said. "The definition of what an agent can do is not the same definition of what the job is that gets hired in the economy." Avinash Vootkuri, a staff data scientist at a top Fortune 500 retailer, said that most enterprise AI agents "absolutely require a babysitter" and, for now, can only work in enterprise settings with tightly bounded autonomy and extensive guardrails. "The stakes are massive," he explained. For example, he described building an agentic system for enterprise cybersecurity where AI agents don't simply trigger alerts and wait for human review, but actively investigate them. Instead of flooding analysts with thousands of warnings, the agents gather evidence in real time -- querying threat-intelligence databases, analyzing behavioral patterns, and filtering out false positives -- before deciding whether a situation warrants escalation. The system relies on tightly bounded autonomy and extensive guardrails, reducing human workload without removing oversight. In cybersecurity, he explained, if the agent gets it wrong, the consequences are immediate and severe. "The AI either blocks legitimate customers (causing massive revenue loss) or it lets a sophisticated threat actor into the network," he said. "It absolutely matters if things go wrong." According to Breanna Whitehead, who runs an AI operations consultancy where she builds AI-powered systems for executives and founders, the industry is in a "trust calibration phase." AI agents can do more than most people let them, but less than the hype suggests. "The real skill isn't building the agent -- it's designing the handoff," she explained. "Most people either over-trust agents and end up cleaning up messes, or they micromanage every output and wonder why AI feels like more work instead of less." The idea, she said, is to design clear handoff points, where something might be fully delegated, another thing might get a quick review, while another task stays just for humans to do. For now, she said, agents are "genuinely excellent" what she called the middle layer of knowledge work -- "the stuff that used to eat 2-3 hours of a smart person's day, like synthesizing meeting notes into action items, drafting follow-up emails in someone's voice, pulling together research briefs, organizing competing priorities into a clear plan." But anything that requires reading a room, navigating ambiguity, or making judgment calls that depend on relationships are not ready for AI agent prime time. "I had a client who wanted to fully automate their investor communications," she said. "The AI could draft beautifully, but it couldn't sense when a funder was losing interest and needed a different approach. The agent drafted the email, but the human had to decide whether to send it." For now, working with AI agents may have less to do with sleeping while they work than with staying half-awake while they do. Tools like OpenClaw can run for hours at a time, but for many early users, that autonomy comes with a new kind of vigilance -- checking logs, reviewing outputs, and stepping in before things go wrong. That dynamic was captured in a recent viral post titled Token Anxiety, in which investor Nikunj Kothari described a friend leaving a party early -- not because he was tired, but because he wanted to get back to his agents. "Nobody questions it anymore," Kothari wrote. "Half the room is thinking the same thing. The other half are probably checking the progress of their agents. At a party." The dream of AI that works while you sleep may be real. But for now, it's still keeping a lot of people awake.

Developer's Honest Assessment of AI at Work Rattles the Official Narrative

A veteran programmer shared his brutally honest opinions about AI's role in the workplace, and it's as much an indictment of the tech as it is of the organizations lazily deploying it. In an X rant that's being praised in online programming circles, the programmer, Dax Raad, said that what's holding back software companies isn't the speed they're able to churn out code, but the quality of their ideas -- an issue AI isn't going to solve, despite the industry's fixation on emphasizing its supposed ability to supercharge productivity. "Your org rarely has good ideas. Ideas being expensive to implement was actually helping," wrote Raad, whose own company OpenAuth sells AI tools. And workers aren't using AI to be ten times more effective, he continued; instead, "they're using it to churn out their tasks with less energy spend." Worse yet, the "two people on your team that actually tried are now flattened by the slop code everyone is producing, they will quit soon." "Even when you produce work faster you're still bottlenecked by bureaucracy and the dozen other realities of shipping something real," Raad concluded. There's some research that backs up Raad's scathing assessment. An ongoing study reported in Harvard Business Review which monitored two hundred employees at a US tech company found that AI was actually intensifying the workers' jobs, instead of reducing their workloads. Using AI to accelerate tasks, it turned out, was a double-edged sword, because it led to "workload creep," forming a vicious cycle in which AI raised expectations for how fast the workers had to churn stuff out, which in turn made them more reliant on AI to keep up with the greater demands. The upshot: worker fatigue, burnout, and lower quality work; not the hallmarks of a thriving organization. Another study documented how AI led to employees passing off low quality "workslop" that masqueraded as good work but in reality required someone else downstream to fix it. On top of slowing everything down, it bred resentment among coworkers, with some admitting that receiving workslop from a colleague lowered their opinion of them. As Raad makes clear, AI is not a cure-all. And even if AI does increase productivity, that productivity can be a mirage. How often are AI models producing shoddy code? And what if that shoddy code goes unnoticed? Maybe, as Raad suggests, ideas being "expensive to implement" was a good thing, because it forced engineers to think about a problem creatively. Not every impulse should be entertained. What's a thousand ideas that were dashed off with an AI instead of a few promising ones that are honed and given time and attention? The former may seem more productive, when it's really a collection of dead ends. Moreover, having employees become dependent on AI hardly seems conducive to rewarding and fostering creativity. As numerous experts have warned, it's another form of cognitive offloading, in which crucial functions of our brain, including critical thinking, are outsourced to a piece of technology. This isn't the line being peddled by tech companies, however. Nvidia CEO Jensen Huang reportedly told his workers they'd be "insane" not to use AI to complete every possible task. Microsoft's AI CEO Mustafa Suleyman claims AI's already so effective that virtually all white collar tasks will be automated within a year and a half. And Microsoft and Google both brag that over a quarter of their code is now AI-generated. But however useful these AI tools may or may not be, they can't work miracles. At the end of the day, it comes down to humans to run a tight ship. "Even when you produce work faster" with AI, Raad said, "you're still bottlenecked by bureaucracy and the dozen other realities of shipping something real."

The CEO of a $1 billion AI unicorn says his peers in Silicon Valley want you to fear for your job, but they're actually first on the chopping block | Fortune

Silicon Valley's artificial intelligence (AI) boom has sparked widespread panic about the future of human labor, a moment summed up by AI executive Matt Shumer's viral essay likening this moment in white-collar work to February 2020, before the pandemic devastated American life. Shumer warned that white-collar workers have to figure out plan B right now, because a Covid-like extinction event is coming for white-collar work. Almost simultaneously, Microsoft's AI chief Mustafa Suleyman gave it 18 months before anyone looking at a computer for a living will be out of work within that timeframe. This was a revival of sorts for the sort of doomsday predictions that marked the first half of 2025 before going ominously silent. Anthropic's Dario Amodei, for instance, predicted that AI would eliminate half of all entry-level white-collar jobs, while Ford CEO Jim Farley said it would wipe out half of white-collar jobs, full-stop. Tanmai Gopal says these dire predictions are a classic case of Silicon Valley self-projection, even narcissism. The co-founder and CEO of PromptQL, a $1 billion-plus Bay Area unicorn that helps companies with AI adoption, told Fortune in a recent interview that the AI doomsday predictions definitely contain a grain of truth while also being massively overstated. "That's 100% what's happening where you have a bunch of ... people who are in the hype cycle." Gopal said his community in the valley is "feeling the awesomeness of this AI" but "we're projecting that into domains that we don't actually understand." "It's like, oh, this is the problem for 7 billion people on the planet, because I'm in Silicon Valley, so I obviously know what's best, right?" Gopal also noted that cynics have a point, with these doomsday predictions occurring right around the time of the next funding multibillion-dollar funding round for many AI start-ups that have yet to go public, offering a clear funding rationale that may not bear out. In general, he added, "Tech people... think like, this affects me. So it's going to affect everyone like that." Actually, Gopal said, that's just not the case. But when it comes to coders, even the senior software engineers, who are exposed to the "awesomeness" of the AI tools now available, he said those people are facing a paradigm shift. Gopal was speaking to Fortune weeks after the "SaaSpocalypse" wiped out $2 trillion in software-as-a-service valuations, with investors realizing, as Bank of America Research recently put it, that AI is a "double-edged sword" and not purely an upside play. It could very easily "cannibalize" many businesses, BofA said, such as software that AI is advanced enough to write itself. Economists have been puzzling over very noisy data over the last year or so, with the U.S. economy largely flatlining in job production while also facing elevated tariff costs and far fewer immigrants entering the workforce. Some AI thought leaders, notably Stanford's Erik Brynjolfsson, looked closely at the data and saw productivity really starting to lift off in 2025. Writing in the Financial Times op-ed, Brynjolfsson noted the latest jobs report revised all job gains for 2025 down to just 181,000, while his own calculation projected productivity of 2.7% for the year, versus the 1.4% average over the past decade. Of course, this lends weight to the AI displacement theory, with even Federal Reserve Governor Michael Barr recently warning that millions could be "essentially unemployable" in the near future. Gopal said it's true that the tech industry has inadvertently automated itself, reaching the era of "baby AGI" (Artificial General Intelligence) specifically for coding. The latest AI models have the judgment and taste of an "average senior software engineer," Gopal said, explaining that standard software engineering heavily relies on converting established business context into technical code and because AI excels at this translation, coding has become the first major domino to fall. "What used to be kind of sometimes considered the epitome... of white collar was like high-grade software engineering," Gopal noted. "That's been all the rage for the last 30 years and I'm excited to see that go." He explained that his excitement stems from the robotic nature of the jobs that robots are already starting to perform and what he's seeing on the frontlines of his company, which helps Fortune 500 companies actually build AI tools and agents that are specialized to their business. "What we've been doing over the last year is ... we've been working exactly at that intersection," Gopal said, and for the most part, he's found that "AI is not useful" because it needs so much business context to be effective. "People keep thinking it's a technical problem," but it's really about the difficult fact that AI can't access business context that lives inside people's heads and hasn't been translated to data -- and may never be. "People are thinking, 'Oh, it's like a semantic layer and a data problem and get your data ready and make it work and whatnot," but the real issue is that data doesn't exist for the most useful information that the AI needs. "Nobody wrote that down. And if nobody wrote that down, you can't train AI on it." Paradoxically for an AI executive, Gopal said that arguably, many businesses exist that AI can never be trained on, "because this is real-life business that moves." Real people who have conversations and continually update a business context will always be one step ahead of the machines, he explained. "Are you going to retrain for that one individual conversation for one day?" he asked, and then retrain on a rolling basis every time your business context changes? Gopal agreed with his interviewer that journalism was an example of a profession that could resist automation, because readers are interested in human insight, deep sourcing and forward-looking analysis, things that AI can't easily reproduce, if ever. He also mentioned salespeople, marketers and operations staff as examples. People in the field who have to make real-time decisions are inherently protected, in his view. Gopal isn't the only executive who recognizes that AI requires human deployment to function. Tatyana Mamut, a former Salesforce and Amazon Web Services executive who now offers AI agent-monitoring purposes through her startup Wayfound.AI, told Fortune that "we need to stop talking about AI like tools. It is not a tool, right? It's not like a hammer." Rather, she argued, it's more like a hammer "that thinks for itself, can design a house, can build a house better than most people who work in the construction industry can build a house." It still needs to be shown the construction plans, though. Regarding business context, Mamut said she thinks "very few" people really understand how to make this work with AI. "You need like real tools and mechanisms to capture that contextual learning." Companies with different brands, different systems and different processes all have different context that need to be captured by AI, she said, predicting that the smart SaaS companies will pivot into this territory. Instead of software-as-a-service, she said expert services will be delivered via agents with proper context capture. Gopal was bearish about how much this context can be captured, estimating that 70% of the effort required to make AI useful relies entirely on unwritten business context that exists only in human heads. "You fundamentally cannot train a system" on this fluid daily reality, Gopal explained, noting that real-life business constantly changes based on individual conversations and human interactions. While AI can automate tasks at the absolute top (coding) and the absolute bottom (physical robotics), the vast middle ground of knowledge work requires human context. Ed Meyercord has been deploying machine learning processes for over a decade at Extreme Networks, a networking company that powers pro football and baseball stadiums and draws in over $1 billion in revenue. He told Fortune in a recent interview that he sees dynamics similar to Gopal's on the operator's side of the table. His teams already use agents to design networks, spot failures before they happen, and even communicate with other agents in systems like ServiceNow, but he is adamant that there is always a human in the loop to review the work when the stakes are critical infrastructure. "A network is critical infrastructure, so we have to be right," Meyercord said. Extreme has built an agentic core into its platform, he added, "but effectively what that's allowed us to do is to be highly, highly accurate." Because accuracy is so paramount, he said, "we always want to have a human in the loop, show all the work that we're doing." Like Gopal, Meyercord said he doesn't believe AI can simply "take our jobs" outright; the role of the human is shifting from doing every task manually to orchestrating agents, gathering the right context, and deciding which problems to point the machines at. He said his job as CEO is, in many ways, to surround himself with specialists "a lot smarter than I am" while using AI as another hyper‑fast teammate rather than a replacement. On the other hand, anything that can be automated is already vulnerable to AI, Gopal said, nodding to the "SaaSpocalypse" in markets that is brutally punishing software-as-a-service stocks, insurance, wealth management and customer service. By the end of the year, he said, this will be even more visible in company valuations, as robots hoover up the work of anything that doesn't require business context. The exciting thing, he added, is what this means for work. This symbiotic relationship between the human worker, who has a business context, and the AI, which can work faster and even smarter but lacks the input, will define the future of white-collar work that Shumer has warned about, according to Gopal. "You have to pick and choose the context and you have to keep capturing the context, right? And I think that's really what the shift is for the average white-collar worker is that they have to understand." Gopal related an anecdote from his team, expressing frustration with a mediocre software engineer now that they have AI coding tools. "We're like, 'Man, like, it's just more expensive to talk to you than it is to do it myself. Like, to explain what I need built on the product takes more time than me just slamming it out of AI on the side.'" The time it takes to talk to a mediocre engineer could be spent managing an AI output instead, he added. He likened this to every employee having a personal technical co-founder by their side at all times, potentially enabling them to produce 20 times as much work. Meyercord agreed, saying that computer-science graduates don't need the same skillset as before, but they will "need a different skillset." He said he's already starting to see new skillsets develop, not necessarily all liberal arts graduates who are deeply trained in critical thinking, but more a sense of "people that are helping us develop." He needs people who can delegate work to AI agents, talk with agents, vet their work, and oversee workflows. It sounds a lot like what Gopal predicted. The job of the human has to evolve to feed the proper inputs to the AI agents that will power the business, Gopal predicted, and he put a name on it. "Our job as humans and people is that we are now context gatherers instead of just workers." Most people have taken this for granted up until now, he said, because they didn't have AI agents to work alongside. "What makes us good at our job, and what gives us promotions, and what makes us more impactful is actually that ability to gather context. That's what makes us good." The only people who genuinely need to fear for their jobs, Gopal warned, are those who are "refusing to grow" and deny this new reality. If everyday workers fail to adopt these tools, they risk handing all economic power to a select few who do understand the technology, potentially creating a dystopian wealth gap. But for those willing to adapt, the future is incredibly bright. "I don't think AI will just come and take our jobs," Gopal said. "That's not even kind of possible". Meyercord said his business is still growing, and he argued that the AI job-loss narrative misses the forest for the trees. "On the one hand, you can do a lot more with less," he said, "or you could do more with the same [number of workers]. Or you could do a lot more with a little more, right?" If you hire the right context gatherers, Meyercord added, you can really grow your business. "It's like, how do you think about what you want to try to accomplish? We want to do a lot more."

No one can agree on whether AI is the next big thing or all hype. Here's why

AI is either your most helpful coworker, a glorified search engine or vastly overrated depending on who you ask. And no one seems to agree on which is right. Tech executives championing AI have long spun the narrative that the tech will revolutionize jobs and bring about a new industrial revolution. Skeptics think it's all marketing hype, while some researchers and executives are sounding the alarm about safety concerns on their way out the door. The discrepancy in how people view AI has perhaps never been so apparent as this past week, after a viral essay from an AI CEO and investor claimed the tech is coming for any job that involves sitting in front of a computer. But there may be a simpler explanation as to why people have taken such divergent stands: People use different types of AI in different ways, yet it's all being referred to in the same way. "There's just a wide spectrum of how much people have been exposed to the technology, how much they've used the technology," said Matt Murphy, a partner at Menlo Ventures who has led investments in AI companies including Anthropic. "And that's also changing pretty rapidly." People who use free AI for basic tasks like making grocery lists and planning vacations are likely only seeing one side of the technology. A report from Menlo Ventures published last June estimated that only 3% of AI users are paid subscribers, although Murphy told CNN he expects that to change quickly. But those who pay get access to another feature: Agents that can handle some work for you rather than just chatbots that craft responses, plus fewer limits on usage. Anthropic's Claude Cowork agent, for example, is only available in the $20-per-month Pro plan and higher. The case is similar for OpenAI's Codex coding agent. It's that type of AI that's fueling concerns about AI's impact on jobs, including the controversial argument Matt Shumer, an investor and former CEO of AI startup, gets at in his viral essay. "I'll tell the AI: 'I want to build this app. Here's what it should do, here's roughly what it should look like. Figure out the user flow, the design, all of it.' And it does. It writes tens of thousands of lines of code," Shumer wrote. He went on to claim that the AI was able to test the app and make decisions concerning taste and judgment. And he surmised that if AI could write code that well, it could begin improving upon itself, as well. (AI researchers accused Shumer of exaggerating the performance of his AI model in 2024. He apologized at the time and told CNN it was the "biggest mistake" of his "professional life" and that he learned through the process). Some experts are skeptical the use cases Shumer outlined are possible even with paid plans, especially since he was vague about which model he used and what type of app the AI built for him. Shumer told CNN he primarily uses OpenAI's GPT-5.3 Codex tool and that he was working on a "medium to high complexity app" for testing purposes. Still, the free version of AI apps doesn't paint the full picture of what the technology is capable of, according to Emily DeJeu, a professor who teaches courses about the use of AI in business at Carnegie Mellon University. She said it would be "misguided" to make assumptions about AI's capabilities based solely on free AI services. Oren Etzioni, professor emeritus at the University of Washington and previous CEO of the Allen Institute for Artificial Intelligence, described the gap between the free and paid tiers of AI like comparing an eager yet inexperienced intern with a seasoned, hard-working intern. Free AI tiers are good at writing summaries and generating content, but users will typically have to pay to conduct deep research or draft sophisticated documents using AI. While free AI can "give you surprisingly good advice" and "engage with you in a surprisingly sophisticated dialog, you wouldn't want to use one of those as your attorney or even as your paralegal," he said. But AI companies are increasingly trickling more advanced features down to the free tier, which is part of why James Landay, cofounder of the Stanford Institute for Human Centered AI, said he doesn't see a big difference between free and paywalled AI. Case in point: Anthropic launched a new model called Sonnet 4.6 on Tuesday that it says will bring performance closer to its more advanced Opus models only available in its paid plans. Software stocks plummeted in early February after AI company Anthropic released a tool tailoring its AI helper specifically for individual industries, like legal and financial analysis. That launch, followed by Shumer's essay, stoked concerns that AI will eventually broadly automate knowledge work the way it's starting to streamline software engineering jobs. Yet there's also growing skepticism about whether AI is living up to these lofty declarations, often made by tech executives with financial interests in the technology's success. Some studies have poured cold water over how capable AI truly is and how quickly it's being adopted. A group of researchers from the Center for AI Safety and Scale AI found last year that leading AI models produced flawed results when tasked with work assignments like visualizing data and coding video games. An organization that tests AI models called Model Evaluation and Threat Research found in July that developers take 19% longer to work on their code when using AI, although that research was based on tools from early 2025. Landay also says the role AI is playing in software development is overstated in the essay. AI is a helpful tool that programmers use to speed up development, but it's still prone to mistakes, and AI models aren't writing themselves. While experts have widely agreed AI will change many industries, AI's proficiency at coding shouldn't be taken as a sign that it'll perform the same way in other professions. "(Coding is) also a logical structure, which is a really good fit for a machine to also be able to test the code and see if it works," he said. "Many people's jobs are not structured in that way."