# My AI Night Shift > Source: > Published: 2026-06-29 11:00:00+00:00 Most mornings now, I’ve got a new routine. Before coffee, before social media, I check my phone to see what got done while I slept. Last Tuesday: a 2,000-word briefing on the risks of helium shortages to Asian semiconductor companies for my consulting clients. The report covered a dozen companies in some detail, touched on second-order effects, and, most importantly, surprised me. There was the nightly security audit of my software platforms—a small bug had been found and fixed. Another message confirmed that my daughter had received her French reading practice. Also completed before dawn: a huge customer segmentation analysis, which simulated 40 people reviewing a proposed new product. Astute readers noticed the passive voice in the previous paragraph. That was intentional. I didn’t complete any of that work. My small staff of [AI](https://www.fastcompany.com/section/artificial-intelligence) agents did. R Mini Arnold (RMA) is an AI agent I set up in February. Why RMA? R is a nod to Isaac Asimov, whose robot names often begin with R, and they serve humans without an agenda. Arnold is the T-800 from Terminator 2: not the killer, but the reprogrammed protector, which is tireless, literal, and utterly loyal. Mini is a tribute to the popular tiny Mac, an elegant desktop box. RMA coordinates several other agents working in parallel: One is an editor, another a researcher/analyst, a third takes an investor’s lens, and a fourth focuses on data. Some evenings up to 10 general coding agents might join them. Together, they handle anything I queue up before I go to bed. RMA determines how best to deploy this squad to accomplish the evening’s work—the security sweeps, the software updates, the tedious technical housekeeping. The customer segmentation report was pretty good, but not perfect. I took 10 minutes to tweak it before I shared it with my managing director. If RMA hadn’t done that work, I’m not sure when we would have tackled it. Hiring an agency would have cost many thousands of dollars and taken too long. Every night, these types of analyses run. They let me close the open loops in my mind—the couldas, wouldas, shouldas that can torment a small-business owner. They help me make decisions more quickly. After only a few weeks with RMA, I had the strange feeling of having caught up with things I’d planned to finish months later. This industriousness has brought a new problem. Silicon Valley investor Tomasz Tunguz calls it the “done list.” And having a done list is an odd feeling. What now? For most people, using AI hasn’t changed much over the past two years. You may have heard colleagues refer to their usage as “fancy Google.” Type something into ChatGPT, get a response, type back. More sophisticated users can ask for a research summary or a spreadsheet. Until very recently, you couldn’t hand anything complex to the AI and walk away. But that changed in November 2025, when Anthropic released Opus 4.5. The new model could reliably follow up to roughly 100 instructions. It was best at writing code, but pretty good at other things. Less than three months later, Anthropic released Opus 4.6, increasing the model’s capacity. A research group called METR has been tracking what it calls “time horizons”—how long and complex a task an AI agent can execute without human intervention. In early 2024, the time horizon was limited to a few minutes; a year later, about 20 to 30 minutes. Soon, agents should be able to handle a daylong task reliably. And within a few years or less, that might well stretch to weeks. This is the AI equivalent of Moore’s law. Moore tracked transistors; METR is tracking time—specifically, how much of a knowledge worker’s day an AI can own without supervision. Call it the Time Horizon law: Where Moore’s law doubled computing power every two years, the Time Horizon law is doubling cognitive reach every four months. And it’s accelerating. To be useful, RMA needs to run on a set of persistent instructions, which can extend to several thousand words and include details and data about me: the name and roles of my team at Exponential View (we research the development of AI in the economy), the major projects from the startups I’ve invested in, my new book, as well as personal details like my love for eclectic EDM. These simple text documents also tell RMA how to behave and how to communicate with me. They define RMA’s “personality,” priorities, and resources. I assign high-consequence tasks to RMA (product development, pricing, customers, my daughter’s revision plans) because if I don’t, I’m not going to care enough to take the time to painstakingly check the agent’s work. To be useful, RMA needs access to the systems I use: email, calendar, code base, project management, and CRM, to name a few. Of course, all of this brings real risks. It can send emails on my behalf, execute code, browse the web. You don’t need to be paranoid to imagine the potential disasters. One of the tools I use, OpenClaw, is insecure and really hard to set up. Unleashed on the world in late January, the software prompted Nvidia CEO Jensen Huang to declare that every company needs an OpenClaw strategy. It’s powerful, but brittle and finicky. One Silicon Valley luminary describes OpenClaw as a Ferrari you have to maintain yourself. RMA breaks at least a couple of times a week. Getting it restarted can take hours. Partly for this reason, RMA lives on its own computer. I can cut the power to it from my iPhone if it goes haywire. I’ve not yet had to do that. The morning review doesn’t involve meticulous verification anymore. There’s simply too much to review. Now I’m validating. Is this roughly in the right direction? Did it move our thinking forward? Does it open a conversation with a client? Validation requires judgment, not deep attention. My judgment muscle doesn’t tire as quickly. But for my human team, that creates a new problem, because I’m now careening through my AI-generated to-do list, prompting a fire hose of detailed requests for my human colleagues. Have we thought about foreign-language editions? Could we build this? Wouldn’t it be better if we worked on that? These aren’t quick emails. They are fully researched documents, with trade-offs and implementation plans. My colleagues have to figure out how to accomplish this work, how to adjust priorities. It occurs to me that maybe I should, you know, just talk to them. This suggests why so many companies see small pockets of individual wins but can’t yet point to AI’s contribution to the bottom line. So much of the critical work of organizations isn’t just execution but coordination. Talking to one another. Agreeing on who does what and when. Deciding who is ultimately responsible for getting the work done. In a 2025 survey, McKinsey found that 88% of organizations have adopted AI in at least one function, but only about a third are restructuring their workflows around it. The AI is ready. The individuals are ready. But the organization itself isn’t. Integrating essential human deliberation into new workflows happening at the speed of AI remains a challenge. Right now, we humans are the bottleneck. In early 2024, Sam Altman predicted a one-person billion-dollar company was imminent, thanks to AI. Matthew Gallagher seems poised to get there with his two-person GLP-1 telehealth startup, Medvi, in 2026. Every enterprise function of Medvi (platform code, website copy, ad creative, customer service) is accomplished via agentic AI. What does that mean for a 20-person company? A 200-person company? I’ve got a team of humans who work for me. But I’ve also got a staff of AI agents doing a team’s work, and it’s only costing me a few thousand dollars a month. Workers will need to know how to direct, monitor, and manage AI agents, and the enterprise will need to invent new workflows. One recent evening, I was lying in bed feeling guilty about not having enough work to give to my agents. I’m not feeling guilty toward them—they are software, after all. I felt guilty about being the kind of guy who has these amazing capacities at my disposal but has run out of useful things to do. That sounds absurd until you experience it. The anxiety wasn’t that the AI couldn’t do enough. It was that I couldn’t give it enough to do. This question of how to tackle the human bottleneck and how to design around it is not going away. As autonomy improves, AI agents will be able to work for days and then weeks at a time on the projects we give them. That will almost certainly create new bottlenecks at the layer where we, the decision-makers, work. Whoever figures out how to do that kind of delegation will do very, very well. For those who don’t, I’m not so sure.