Why Agents Still Need Humans

Today on the AI Daily Brief: The next wave of human-agent collaboration. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors: Robots and Pencils, ZenCoder, Assembly, and Superintelligence. To get an ad-free version of the show, go to patreon.com/AIDailyBrief, Or you can subscribe on Apple Podcasts. If you want to learn more about sponsoring the show or really know anything else about the show, you can go to aidailybrief.ai. One thing I would point you to is the newsletter, which is back. If you're ever wondering where you can find the links to all of the articles and quotes and tweets and things that I reference, the newsletter is going to be your best bet for that. Now today we are doing a long read slash big think episode, and we're getting at a theme that's at the core of AI operations this year. Obviously 2026 has been all about agents actually becoming real, and they became real because of a combination of the model advancements at the end of last year, as well as the greater focus on harnesses, i.e., the interfaces through which we interact with agents. Through the combination from January till now, the way that we use AI is no longer sit there, prompt it, wait for an answer, and go off and do the rest of our work. Instead, increasingly, it is about spinning up or managing agents that go out and produce things on our behalf. Agents that can use code to build things or solve problems, even when we're not coders ourselves. And of course, the implications of this have been massive. Business models are shifting as companies are no longer able to subsidize the biggest power users of AI, who can consume hundreds of millions or even billions of tokens themselves individually in a single month. Indeed, more broadly, we are starting to live inside a world of token shortage, where the total amount of AI that would be consumed if it could is higher than the amount of AI that is available, thanks to constraints of compute. Throughout the last few weeks, we've been talking about some of these big implications, but one that we haven't mentioned for a little while now is what it means for the patterns in how we work. At the beginning of the year, as Open Claw excitement raged, it was all about Mac Minis and even for some Mac Studios running 24/7 agents doing everything you could possibly imagine, not only automating your existing world of work, but uncovering new things that were never possible before. And what was interesting in all of this is that both the promise and the fear of AI, the promise of AI that would reduce how long it took to do your work so you could go enjoy more leisure time, and the fear of AI that would negate your value as a worker, were both very far away from the lived reality of the most advanced users. In fact, instead of finishing your workday at 3:00 PM, the more common challenge was people having to force themselves to go to bed at 3:00 AM, tearing themselves away from the next thing they could accomplish, which was always just sitting there waiting for them. Now, I discussed this phenomenon in my episode, Why Agents Make Every Job a Startup. In that episode, I introduced the concept of the infinite backlog and argued that simply put, agents make it feel like for the first time we have beaten the end boss of time. Even in the assisted AI paradigm, there was a reasonable end to your work because you, as the user of AI, simply couldn't do anymore. But agents aren't you doing a thing. Agents don't get tired. They don't have to stop. The only reason that an agent isn't working is if you haven't given it something to do, which means that it no longer feels like there is an actual end. There was always just work that you didn't give the agent. And as it turns out, the amount of work to be done is not actually bounded. There is always something next. This is what I referred to as the infinite backlog. What's amazing about agents is that they can do more of the infinite backlog than was ever possible before. A single person can go deeper into that infinite backlog than was literally ever possible, which is why it feels so incredible to build things with agents that you had never dreamed of. At the same time, agents make it feel like you should be able to do the entire infinite backlog and that anything your agents are not doing is because you haven't given them the tools to do it. This is a very particular and new type of overwhelm that was not on most people's radars when it came to the implications of AI for work. Now, one of the most interesting companies at the forefront of experimentation with AI is Every. Every is part publication, part product company, part consultancy, and really walks the walk when it comes to experimenting with how to run an AI-native company. Every CEO Dan Shipper recently wrote an essay that is effectively his version and his exploration of this same phenomenon that I was looking at with the infinite backlog and the Agents Make Every Job a Startup episode. He called his After Automation. And I wanna read a few excerpts right now. Dan writes, "There is a paradox at the heart of AI. At Every, we've automated everything we can. We use Codex and Claude Code across coding, writing, design, customer service, and more. We alpha test all of the new models from OpenAI, Anthropic, and Google before they come out. We are riding the exponential boom in model intelligence and automation as far and as fast as possible. And yet it seems like for us, there's more human work to do than ever. We're a team of almost 30 people, and we haven't fired all of our employees in favor of agents. We haven't ditched SaaS products in favor of Vivecoded apps. We still hire humans to do customer service with a lot of agent assistance, and we still hire human writers and editors and engineers. Our work does look completely different than it used to though. We don't write code by hand anymore. If you @mention someone in our Slack, it's a toss-up whether you're talking to a human or an agent. Managers are committing code like ICs and engineers are talking directly to customers. For the last several weeks, AI has responded to 95% of my work emails. In short, the future looks weird, but also familiar. The familiarity is surprising because one thing CEOs, knowledge workers, and investors seem to agree on is that AI is a threat to jobs, the economy, safety, and human meaning. Anthropic CEO Dario Amadei warns that AI could wipe out up to half of all entry-level white-collar jobs. Meta just laid off 8,000 people and is installing software on US employees' computers to capture mouse movements, clicks, and keystrokes. For a higher-quality source of AI training data on advanced knowledge work. Even Citadel's Ken Griffin seems shaken, saying recently, "These are not mid-tier white-collar jobs. These are extraordinarily high-skilled jobs being, I'm going to pick a word, automated by Agentic AI." Dan then points out that all the benchmarks seem to validate this set of capabilities. It seems, he continues, that we are on the cusp of an AI smarter than any human, with the autonomy to work for almost a full day at a time. And yet, The paradox remains. If you talk to anyone in the AI industry or to early adopters outside of it, you'll hear the same thing we've noticed internally: there's more work to do than ever. The big question, within the industry and without, is, is this just a temporary state of affairs? Will the next model drop be the one to replace everyone? We watch the benchmarks and sweat, wondering if there's a tipping point around the corner where all of the jobs go away. There's no tipping point coming where things flip and the jobs are gone. The new reality is the opposite. The more we automate, the more expert human work there is to do. Here's why. AI commoditizes the residue of human expertise, whatever can be made explicit enough to train on. That collapses the value of default model output and creates demand for what's different. And demand for what's different is demand for human experts, even as we approach artificial general intelligence. Moving down, Dan discusses the two modes of working with agents. The first, he writes, is the one the AI discourse predicted pretty well: agents as employees. These are agents you delegate work to. Some are agents that live in Slack, have names and jobs, and can be tagged when you want them to do something. Some are agents embedded in an ongoing workflow, like customer service, acting as always-on gatekeepers for repetitive tasks. The second mode is stranger, and in my experience, more important. It is human-agent collaboration in tools like Codex, Claude Code, and Claude Cowork. These are not just places where you hand off work. They're becoming operating systems for the work itself, where you and multiple agents use the same computer at the same time to do highly complex original work that can't be done by an asynchronous agent. In both of these modes, you can use AI to automate and delegate much of your work, and both of these modes require you or another human in order to work well. Dan then talks about the different types of employees that they have running around at Every. "Agent employees," he writes, "are given a job and go off to produce an answer, an action, a report, a draft, a triage decision without you in the loop." These include coworker agents, i.e., agents you can tag in Slack and ask to do work. One example is Andy, their editorial team's coworker agent. Andy collects nuggets, which are good ideas for stories pulled from internal Slack, then turns them into digests and first-pass takes that writers then use to compile the daily newsletter. Embedded agents are agents that live inside a particular product's workflow. Dan writes that these agents are less flexible but can be powerful for helping with repetitive tasks. He points to Finn, which is an agent embedded in their customer service platform, formerly Intercom presumably, that handles a lot of their support load through chat and email. Across both forms, coworker and embedded, Dan continues, the pattern is the same. Employee agents take over more of the stable, repeatable, well-framed layer of work, but there is a lot of work that still requires a human being in the loop. We found over and over that for any kind of complex task, The best way to get great work is to have an AI and a human going back and forth in the same workspace. This is what Codex, Claude Code, and Cowork are for. They allow you to spin up and delegate work to one or more agents across multiple chat threads. These agents have access to your computer and all of your sources of data. You can see each task the agent is doing and thinking about and can interrupt at any time. And you're responsible for managing the agents at the start and the end of each one of their tasks, making sure it's done well and finding the next piece of work to do. One of Dan's employees at Everly calls this the human sandwich. With humans as the bread on either sides of AI's work. On one side, human sets the frame of what they're trying to do and what counts as good. The AI then collapses the task into draft, searches code, summarizing, and comparison. And then the human judges and extends the work. Is this good? Where does it belong? What should happen next? Now, what Dan points out next, and something I highly encourage you to go check out in the original piece, which is of course gonna be linked in the show notes, is that while coding is an obvious example of this work pattern, it's coming to the rest of knowledge work as well. One thing I keep seeing in enterprise AI: companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends. The teams actually shipping, they've picked a lane and they move fast. That's one of the reasons I like today's sponsor, Robots and Pencils. They've gone all in on AWS. They're an advanced tier and AWS pattern partner, and they ship production AI coworkers in 45 days. That's led to them doing some of the more interesting work I've seen on AI coworkers. And by that, I'm not talking about chatbots. I'm talking about actual agentic systems that sit inside a business architecture and do real work. That kind of focus matters if you're an enterprise leader trying to get something real into production or an AWS rep trying to move a customer from interested to deployed. Request an AI briefing at robotsandpencils.com. One conversation with Robots and Pencils and you'll know. It is a truth universally acknowledged that if your enterprise AI strategy is trying to buy the right AI tools, you don't have an enterprise AI strategy. Enterprise AI strategy. Turns out that AI adoption is complex. It involves not only use cases, but systems integration, data foundations, outcome tracking, people and skills, and governance. My company, Super Intelligent, provides voice agent-driven assessments that map your organizational maturity against industry benchmarks against all of these dimensions. If you want to find out more about how that works, go to besuper.ai, and when you fill out the get started form, mention maturity maps. Again, that's bsuper.ai. So coding agents are basically solved at this point—they're incredible at writing code. But here's the thing nobody talks about: coding is maybe a quarter of an engineer's actual day. The rest is standups, stakeholder updates, meeting prep, chasing context across 6 different tools. And it's not just engineers—sales spends more time assembling proposals than selling, finance is manually chasing subscription requests, marketing finds out what shipped 2 weeks after it merged. Zencoder just launched ZenFlow Work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Jira, Gmail, Google Docs, Linear, Calendar, Notion. It runs goal-driven workflows that actually finish. Your standup brief is written before you sit down. Review cycle coming up? It pulls 6 months of tickets and writes the prep doc. Now you might be thinking, didn't OpenClaw try to do this? It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve. ZenCoder took a different approach. SOC 2 Type 2 certified, curated integrations, tighter security perimeter. Enterprise-grade from day one. Model agnostic and works from Slack or Telegram. Try it at zenflow.free. You know AssemblyAI for having the most accurate streaming speech-to-text out there, but they just went a step further and launched a full voice agent API. The idea is simple: one connection and they handle everything. The listening, the thinking, the speaking. You just stream audio in and get your agent's voice response back. We're talking about things like outbound sales calls that actually qualify leads, customer support that handles complex requests without a script, scheduling agents that sound like a human assistant, and you can build one in 5 minutes with one API. And importantly, their streaming model is the best at catching all the stuff that breaks on other voice agents, things like phone numbers, emails, names, and medical terms. And for those of you who are still in experimentation mode, there are no contracts and unlimited concurrency, so you can actually test it out without any friction. Head to assemblyai.com/brief and try the live voice agent demo right there on the site. No signup needed. He describes how he uses it for writing an email, for example, and what he lands on and what the human sandwich implies is that agents need humans in order for the work to work. Now, interestingly, I noticed a couple weeks ago that Evry had also shifted their philosophy of agents in a pretty dramatic way. Initially, around the first blush of Open Claw excitement, Every had basically every employee spin up their own AI agent who was a replica of themselves. The problem they found very quickly was what happened when everyone had their own agent. In another essay reflecting on their first set of experiments, they wrote, "Every time an agent broke, the person it belonged to had to fix it for themselves. Even with a stable harness, agents require maintenance to perform. This was great for someone who likes tinkering—the maintenance and back and forth are part of the appeal. For every tinkerer, however, there are a lot of people who want the benefits of an agent without the obligation of having to manage and mend it. What they discovered, they wrote, is that rather than agents as extensions of their creators, a more successful model is agents as coworkers who reliably perform parts of many different people's jobs. This, among other things, takes the maintenance burden off of the individual. They continue, imagine a shared analytics agent. Everyone on the team uses it for metrics-based work. And when its capabilities need to expand, one person updates the agent's skills and the whole team benefits. In the personal agent version of the same scenario, that same update has to happen across 10 different agents. Team-based agents also solve a continuity problem. A personal agent's value is tied to whomever trained it and disappears if that employee leaves. A team agent with defined capabilities retains company context and knowledge, acting more like a project manager, sales lead, or chief of staff than a private assistant. Now, this maintenance was part of the reason, going back to the essay we started on, why agents were creating more work for humans. But Dan points out that there is a second reason as well. Continuing with the After Automation essay, Dan writes, "If you look at AI's exponential trajectory over the last few years, and think about how its architecture works and where its powers come from, you'll see clear feedback loops that create more human work. AI makes yesterday's human competence cheap. Language models are trained on the visible residue of human competence." Code, prose, images, support tickets, product specs, and more. They take all of it—the exhaustive successfully completed tasks—and package it in a form that's available to anyone cheaply. The net effect is that skills that used to be rare—coding a pull request, making a YouTube thumbnail, writing a newsletter—are now broadly available to almost anyone. Cheap competence gets rapidly adopted. When the cost goes down for something previously rare, supply suddenly goes way up. At Every, we see this all the time. Operations and customer service people are writing code and issuing pull requests. Marketers are making YouTube thumbnails. Engineers and product people are writing drafts of articles, guides, and landing pages where they never would have before. Abundance creates sameness. Old expertise becomes commoditized. Because everyone has access to the same models, and the models are all based on yesterday's competence, by default, the models end up creating work that ranges from "a decent start" to "it's just plain slop." Slop is not any one particular mistake. It's not the use of em dashes or a certain sentence rhythm or purple accents on a landing page. Slop is visible sameness repeated ad nauseam. It is what gets produced by default when humans in many different circumstances use the same tool, train on the same corpus without thinking too hard. It's what happens when everyone has access to an expert who has the same default tendencies. An abundance of sameness rapidly becomes a commodity. But sameness creates a demand for difference. Humans, Dan points out, very quickly spot this sameness and want something better. Demand for difference, Dan argues, is new demand for experts. Because of the architecture of language models and their broad distribution to everyone on the planet, Rare and valuable work must come from a human. The current generation of models only knows about work that has been done. Humans know about what needs to be done right now at this moment. This is the paradox we started with: making expert work cheaper does not simply replace experts. It creates more situations where expert judgment is needed. In response, human experts move in two directions at once. Some use AI to build systems that absorb and leverage the flood of new work. And some use AI to do bigger, more interesting work than they could have done without it. This is why, in practice, AI does not eliminate expert human knowledge work. It dramatically increases the volume of work being done, and none of that work is differentiated or valuable unless a human being is involved. Now, the next section of the essay is Dan dealing with the obvious objection of, but once we get to AGI, doesn't AI do all of that expert stuff and planning stuff as well? In other words, even if AI creates new jobs because of new opportunities, doesn't even better AI just do those jobs too? I'm going to leave that section for you to read, as it's an interesting meditation on the nature of benchmarks, and an argument that from here on out, we'll always effectively be in this race with AI, where ultimately it still is waiting for us to tell it what the next most important thing to do is, even if it's helping us make those decisions. If you're interested in a bit more tangible of an answer, check out my episode on the new jobs AI will create. In it, I introduce a framework I call the human premium, which are 7 categories of value that don't transfer when you remove the human. Basically where even if an AI can do the thing, there are reasons that humans will not want it to do the thing. But where I want to zoom out to is the fundamental insight and observation that based now on the nearly half year of everyone working deeply with agent collaborators in the most AI-native type of setting that you can find, the patterns for how we work with agents are changing, but it is very much still a pattern of us working with agents. So what is actually shifting about those patterns? Well, one example, the one that we just heard from Evry, is that instead of every individual having their own agent that is a digital embodiment of themselves, they are now experimenting with more unit agents that multiple people whose work interfaces with each other's all rely on that same agent, which has all sorts of benefits in synchronicity, maintenance, and more. This is then a pattern that might be interesting to explore for your own organization. Instead of everyone independently having their own open clause or agents you've spun up in Claude code, thinking about where the Venn diagrams of people's work overlaps and asking whether there are agents that if they lived in that overlap would help even more. Another area where you're seeing a maturation and change in the work pattern itself is around the first set of early adopters for OpenClaw. Matt Schumer recently wrote, "Just wiped the Mac Mini I had set up for OpenClaw. I'm turning it into an always-on dev box to use with Codex Mobile. Have a feeling this is going to be amazing." So with OpenClaw, the pattern was one of the minimal possible interaction with your agent. You put it on a Mac Mini to give it access to a set of tools and information, and then you interface with it through something like Telegram. It used the paradigm of heartbeats, which are basically recurring timed reminders to itself to make sure that it was continuously progressing against the goals that you had set for it. But if you take what Dan Shipper from Evry said as exemplary, many people have found that that level of autonomy for agents wasn't actually accomplishing the things that they wanted. One thing that that level of autonomy was good for was absolutely burning tokens. And as we move into this token shortage era, obviously there are actual monetary costs with that level of autonomy as well. Matt's not the only one though, whose work patterns are shifting around agents and harnesses like Claude Code and Codex. Nick Baumann from OpenAI recently tweeted, my laptop has become a satellite device since I started using Codex from my phone. And my Mac Mini has become the home. It's clunky, but the end state feels more like how we're going to be working in the near future. I'm currently running the Codex app on two devices, my MacBook and my Mac Mini. My laptop isn't reliably connected to Wi-Fi enough, so I keep a Mac Mini on my desk that is always connected. When I kick off new threads from my phone, because remember, a full-featured Codex is now in the ChatGPT app, Nick continues, I start them on the Mac Mini. When I'm working from my desk, I run them there too. The cool part is that I've added my MacBook and Mac Mini as connected devices to each other. That means I can start and resume threads from either device. So if I'm in a meeting but want to continue a thread on my laptop that was started on my Mac Mini, I can do that. What this means: I have an always-on codex that is accessible from my phone, with its own dev environment. All threads are always accessible from any of the 3 devices, and I can run heartbeat threads that stay on 24/7. It's a little makeshift today, but the shape of it feels very real to me. Codex is no longer tied to whichever computer happens to be open in front of me. It starts to feel like something I can stay connected to across whatever device I'm using. Okay, so zooming out again, we've got these early experiments in autonomy with OpenClaw that maybe concluded that the managerial burden of that autonomy wasn't the best fit for that particular harness. Meanwhile, these work operating systems in Codex and Claude Code feel a little bit closer to the right way to manage the correct level of autonomy for these agents as they currently exist. And with advances of the UX in these harnesses, specifically features that make them more accessible from different devices and on the go, they become less and less reliant on any given device and more nimble and semi-synchronous. Now, if you listen to my episode from earlier this week about how to get the most out of Codex, the OpenAI author that inspired the piece, Jason Liu, was nominally giving 9 tips for how he gets the most out of Codex. But when you take a step back, they pretty much all come back to how to better parallel process and live in a state of semi-synchronicity with your agents instead of being stuck in some turn-based paradigm. In other words, we don't want the purely turn-based paradigm of assisted AI where you give a prompt, you sit around waiting for its response, you review the output, and then you give the next prompt. But we also need more ability to manage than the mega autonomy of something like OpenClaw that's mostly just using Heartbeats to run itself with you checking in via Telegram. A lot of what people are then experimenting with now is how to use harnesses for some middle space where there's less latency between the instructions and guidance that you need to give the AI and the way that agents can go do that work. So two experiments to consider, one on a personal level, one on an organization level. On the personal level, go check out my episode on 9 Tips for Getting the Most Out of Codex and start to think in terms of reduced latency. How do you use steering features and voice-based input to compress the space between AI doing the work and you guiding the AI such that you're working more synchronously with your agents. Then on a more organization or group level, look at different people's jobs and try to map the overlap—what I called the shared space in the Venn diagrams before—and then instead of having each person ask about what an agent could do for them individually, see what types of tasks live in that shared space and what agents could do for those. That seems to be where AI-native companies like Everly are getting a lot of their current value. And then finally, For those who are trying to mentally work your way through the AI tomb cycle that I introduced this week, especially those of you who are in this real-world recalibration moment where you're taking the evidence of what we're actually experiencing and applying that to thinking about where AI and agents are going to go, I think it's worth taking more seriously the idea that no matter how much of today's work they can do, their net impact on employment is going to be to expand it in proportion to all of the new things that we can do now that we never could have before. And it's not just me that's saying this. Gartner is starting to argue strenuously that even if we see short-term AI layoffs beginning in 2028, they believe and argue that AI is going to create more jobs than it eliminates. And I believe we're even starting to see the very earliest indications of this mindset showing up in markets as well. It might appear like every time there's layoffs, stocks for the companies doing the layoff crank, but I think that there's evidence that increasingly what markets are looking for is not AI efficiency, but AI-related growth. Take the example of Atlassian. In March, they announced 10% layoffs, and that announcement actually marked the end of a mini recovery over the previous 2 weeks. The stock actually headed into a year-to-date low in mid-April. Then, however, Atlassian reported 29% earnings growth for Q1, anchored around strong sales for AI-enhanced products, and that sent the stock soaring 29% that evening. Market analyst Dan Ives recently said, My biggest concern is tech companies tripping over their own shoelaces, talking about job cuts, not reading the room, saying that their technology is going to wipe out jobs for young people. You do that, you just shot yourself in the foot. And when the podcaster he was talking to pushed him, asking if that was just a market narrative thing, or whether people like Anthropic's Zaryll Ahmed were wrong in their predictions of job loss, Ives continued, 100%. What's going to separate companies? LLMs are going to get commodified. What separates companies is the people. It's the engineering, it's the marketing. It would obviously be going too far to say that the whole market is shifting to think the way that Dan thinks. But I think that increasingly the burden of evidence is going to suggest that the companies that are crushing with AI are going to be the ones that A, are investing in their team's capabilities to use and manage agents, that B, recognize that growth, not just efficiency, is what will lead to long-term success, and C, that agents are not a get out of budget jail free card but one of the best investment opportunities that companies have ever had. Sum it all up and I think we are on the cusp of the next wave of human-agent collaboration, and it's gonna be a good one. For now, that's gonna do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.