高速化するAIエージェントと減る人間の同僚：エージェント管理者になる未来（再放送）

I am the only full-time employee at Everyday AI. Sure, I have a few contractors helping out, but I'm the only full-time human responsible for creating this daily podcast that reaches millions of people a year. Others are kind of always shocked to hear that, as most of the other top 25 tech podcasts have large teams to produce the same amount that we produce. But do you know what I do all day? I just manage agents. I mean, literally right now, I have Codex running with at least 50 sub-agents active. I have Codex controlling CloudCode, and I have scheduled agents running in Google Gemini, CloudCodeWork, Copilot, and others. Maybe I'm a little overly productive from a single human to output perspective, but it's about to get worse, or maybe better, depending on your perspective. That's because as soon as today or Friday at the latest, we'll have a new open AI model powered by Cerebrus that runs up to 20 times faster than their current models have the same capabilities.

And that got me thinking, what the heck is the future of work going to be like? I mean, I'm already agent-pilled and producing more content and hopefully value than much larger teams with many more humans. So what happens when all of my agents are 20 times faster? I'm not sure yet, but I wanted to kind of opine about that out loud on today's show and explore this weird convergence that we're in now. Because not only are agents producing economically valuable work end-to-end at a higher quality bar than expert humans, but now we are entering a new phase where those AI agents might be working 20 times faster. And I don't know if anyone truly understands what this convergence means for the future of work, but let's unpack it and see what we might come up with anyways on today's episode of Everyday AI. So here is the big picture. What the heck is going to happen when millions of professionals like me and you, business leaders that are on the cutting edge of AI?

What's going to happen when we just become managers of faster agents and they become more powerful? Right now, AI is fast and parallel enough that one person can direct dozens of agents a day. That's pretty much all I do all day. But OpenAI, as well we are seeing, maybe by the time you're listening to this, it will already be released. But at least as of now, you know, the rumors are saying that we're going to be getting a new OpenAI model powered by Cerebras. That's essentially this really fast chip. And essentially, let's say you put, you know, a very, you know, smart agent out there and it thinks and, you know, normally it takes five minutes and you get a great output. Well, what happens when that five minutes becomes 30 seconds? And normally you go, I don't know, manage other agents or do other tasks during that time that you're waiting. What happens when things are just instantly done? And the upside here is real, right, with workers reporting more autonomy and more outputs.

But the risk is maybe a workday spent managing machines while human trust and mentorship quietly just thins out. So on today's show, here's what you're going to learn. You're going to know why employees are increasingly asking agents before managers, specialists, or awkward conversations. You're going to understand how faster inference turns solo work into supervising parallel machine teams. You're going to know how winning companies can empower human judgment while still automating most of the manual knowledge work. And you're going to leave the show with the three human handoffs to protect before your workflow quietly deletes them. All right. Welcome to Everyday AI. If you're new here, my name is Jordan Wilson. We do this every day. It's for you. It's your daily unedited, unscripted live stream podcast and free daily newsletter helping business leaders like you and me keep up with drinking from the fire hose of new AI updates. I turn that fire hose down. I serve you nice glass of water and say, here, this is what matters today.

Go take this information and share it with your team and become really smart in AI. So that's what we do here. Hopefully the podcast is helpful. If you didn't know, we recap it all in our daily newsletter. So make sure that you go check out our newsletter for that and all of the other AI news. All right. So when did this start happening? I'm not sure. As someone that follows this every single day, this kind of convergence now that is maybe going to start today for people on the edge of the edge. But I think the rest of the enterprise will start to feel this in maybe six to nine months. But we are now going to be at this convergence very quickly where we've had these extremely capable agents now, for the most part, running on desktop. So whether you're using Codex, ChatGPT Work, Cloud Desktop, something else, right? But I'd say those and maybe cursor are the big players, at least when it comes to people, number of people using them. I don't think Google is there just yet with, you know, anti-gravity or their Gemini desktop.

Microsoft CEO Satya Nadella did mention again yesterday their super app coming out. So I do know that most of our many of our listeners out there are using Microsoft Windows and Copilot. So you'll be living in this world soon, too, if you haven't already, maybe with, you know, co-pilot, co-work. But we're at this place now where we have these agents that if you know what you're doing, they will literally on demand output the same level, the same asset, the same deliverable as expert humans. We see this in important benchmarks like GDPVal that a human and a large language model get the same inputs, and then they create outputs. They're judged blindly by subject matter experts. And we're at the point now where AI models win or tie about 90% of the time. So AI models or AI agents can actually do the same work that humans do and create these assets or artifacts or deliverables, which is cool and all right but you know i'd say the downside no it's not even technically a

downside right i always tell people be a little patient but you have to number one know what you're doing number two you have to understand kind of agent guardrails in context engineering which those are kind of basics uh tables uh table stakes now but number three is a little bit of time right and that little bit of time is now what could be changing today whereas before you know I have agents that work for many hours, but I'd say my average agent task maybe lasts about five minutes. So for the most part, all day, I'm just flipping between agents, right? And one of the big things I like doing is doing that on chat GPT voice mode. But what happens then when that five minutes becomes 30 seconds, right? I think of so much of what I actually accomplish day to day. It's when I'm waiting for agents, I go check in on other agents. So then at least, you know, by the time the day is done, right, in theory, I've had time to check in and look at those artifacts and look at those outputs and, you know, manually approve them.

So what happens when things go faster? Well, one thing is I think where we're going to start is workers are just going to start asking agents now, right? Because not only is this convergence of these two things happening faster and smarter agents but also that takes away I think ultimately human connection And I think where you know people for the last 20 years right maybe I been guilty of this managing people in the past and maybe you've said this or it's happened to you, but, you know, saying, hey, did you Google that? Right now, one of the things is going to be like, hey, did you ask our agents on that? Which isn't necessarily a bad thing in theory, right? But it's going to take away human connection, right? Even for me, right? I started the show by saying I'm the only employee at every day, the only full-time employee at every day. I'm not saying that's a good thing or a bad thing, but I'm not talking to a lot of humans all day, right?

For the most part, I'm talking with agents. Yeah. Again, pros and cons, good and bad. I'm able to produce a lot more, but maybe I'm getting fewer outside perspectives that aren't tied to my own custom instructions and anemic data that I'm feeding these models. But right now, studies show, so Adobe just came out with an interesting study this month that showed that 68% of people ask agents obvious questions and only 4% of those people would ask their managers. So people are like more than 15 times, 16 times more likely to ask an agent something than ask their manager. Again, not saying that's a bad thing, but what does that do to human connection? And workers chose the agent because it answers instantly and never makes them feel unprepared or dumb. And then once that agent answers first, workers kind of stopped routing the work to specialists too because then they build more trust.

and they're like, okay, well, I was going to ask my manager on this, but the agent did great. And I was going to take this over to the product team, but I asked our agent and that did great too, right? So, you know, you start to be able to build faster in silos versus maybe building more intentionally at a slower pace, but in a collaborative fashion. So that's the other thing is now all of a sudden we're getting a lot more skills that we didn't have before, you know, and we're able to attempt work that we couldn't previously because we have a new set of skills. You know, kind of the best way for me to look at this, right, way back in the day, you know, I used to have a marketing and advertising agency, right? You know, kind of a big part of my background is, you know, I was a writer, I was a journalist before. And, you know, as a marketer, sometimes you have to wear many hats. But at a bigger company, a lot of times it's you have people who are specialists, right? You have people who just maybe write Google ads. And that's all they've done for, you know, 20 years.

You know, now it's getting to the point where one marketer, maybe they can't do it at an A rating, but they could maybe do 20 jobs that have B rating. Whereas before, maybe they only had one specialty. Right. So now as we get collectively better in the enterprise at sharing our our skill sets, recording and distributing kind of our nuanced understanding and kind of internal IP, so to speak, and give that to others via shared agents or shared skills. All of a sudden, individual humans don't always need their manager's approval or even a specialist to help complete the work. And every handoff, those handoffs, the human handoff that we're starting to decrease, it deletes a human conversation. It deletes mentorship. right handoff is kind of once something is carried through explanation correction introduction in a

moment of shared context so it was perplies i i believe that's how it's pronounced perplies may 2026 study that found that 63 percent of people use ai to avoid a difficult workplace conversation right um and and that's you know that one skipped conversation multiplies um And I think you also start to get internal project drift there. But I think so much of even I'm thinking of where I am today, right? Probably the reason maybe why I've been able to have whatever level of success that I have with this weird little podcast thing is, you know, I had a lot of great mentors throughout the years. I've been able to learn from very smart people. And sometimes that comes from conversations when things go awry, you know, or when things go great and having that be a teachable moment. So I'm just thinking, you know, yes, there's, you know, working with agents, you can get all those things and build them in. But, you know, agents by default are sometimes overly sycophantic, you know, and then I think you can almost develop a certain AI psychosis, you know, in the good way or bad way versus, well, when you had a team of humans.

and collaborating more with humans, I don't think that that was as big of a problem, right? Working in these silos that, yes, maybe you are, you know, two, three, four, five times more productive in terms of the business value that you're able to create, but at what cost? So the translation here is, well, we have more collective skills that we didn't have before, but, and we also have fewer productivity roadblocks, right? The two things that I just covered. And, well, that gives us all the reason to be more productive. And, well, things are getting faster now, too. That's because today, maybe tomorrow, but OpenAI said in July that they're going to release the new GBT-56 Sol on the Cerebris chip, which is up to 750 tokens per second. So if you don't speak tokens, that's about 20 times faster than the current models, right? So if you go into, you know, codecs like I do and, you know, you turn on, you know,

GPT-5-6 sold and you put it on, you know, extra high or high or ultra or something like that, even if you're throwing a hard task, you know, you might only have to wait five, six, seven minutes. Now, all of a sudden, that's going to be a matter of seconds. So the value there is the throughput. It's with the research, the drafting, analysis and review all running simultaneously. So a slow agent is a tool that you wait on. And I don't think that's necessarily a bad thing because what I have always done when I'm waiting on agents, I'm usually reading the chain of thought. Right. If in that five minutes, I'm literally looking at the chain of thought. I'm making sure that it's calling the right tools. I'm making sure it's pulling in all the correct dynamic data because it's like, what else am I going to do? Well, yeah, probably multitask with some some different agents and go in and check on my blue lights that, you know, require my input. But ultimately, I'm spending more time being the expert driven loop. Right. So I think that this is actually could be a big problem. Right. When AI is almost too fast, even the people that are maybe trying to be responsible.

Right. All of a sudden, everyone knows, you know, in a year or so, everyone's going to know, oh, the AI is going to be way more powerful and it's going to be even faster. Whereas right now, if you're using the right way, in theory, right, you're getting more done. But it's almost like you're going slower, but way faster. You know, it's like you're waiting, but you're accomplishing more. Right. So instead of working on how we would traditionally do it, you know, in pre-AI days, you know, me one human I would work on one project across these 20 little steps Now conversely I working on 20 little projects And my only one step is just checking in on what the agents are doing Right But you know usually I stagger it in a way where I can understand kind of, you know, prioritize explainability and, you know, making sure the guardrails that I said are being followed, making sure my agents are crashing all of these things. But what happens when this is done? Am I still going to go and do that? I hope so. But when things

are done faster and when the models are smarter, won't my human nature just be to be like, well, I should go grow the business in a new way. I should go get more and more things done. Right. A fast just team just now needs more management. So directing agents is just middle management without the manager title. Right. I do see the future where most, you know, most knowledge, you know, subject matter experts. Most people, if you're a knowledge worker, you sit in front of a computer and you create value for your business, which is what many of us do. I think people, for the most part, you're going to be like a middle manager, right? Of just agents. I like, I actually think that in theory, the middle management layer is going to get bigger and thicker. I think there's going to be fewer people at the entry level. I think there's going to be fewer people in the, you know, the VP, C-suite director level. I think for the most part,

you know, your average knowledge worker, that middle management tier, right, when we think it's not really needed, well, I think, you know, if we look at middle management today in an AI native world, it's not needed. But I think just everyone else is going to become kind of that middle layer, right? People are going to be managing more agents, not people. And yes, There's still going to be, you know, humans reporting up the food chain. But I think for the most part, the overwhelming majority of work is actually going to get done in that middle management layer, where traditionally that's not where the work gets done. Right. They're just managing people. So it's I think that's ultimately what's going to happen. But I think it's normal in AI native companies for decision makers to just publish an agent to help to speed up productivity. Right. You've seen these stories where literally, you know, a CEO or a VP or, you know, someone that manages a big department at a huge company is just literally unloading their entire brain and decision making process and all their documents and all of these things into an agent.

Right. Let's say it's it's, you know, marketing Jane at a Fortune 500 company. You know, Jane just dumps her brain, her decision making process, everything into an agent. And, you know, 500 people in marketing. Well, they can just get answers straight from the Jane agent. They don't necessarily have to wait. So the work, it just becomes faster, but it becomes four verbs. I think you decide what matters, you delegate, you inspect the exceptions and then you integrate the results. right so but those things can happen now like agents can in parallel and you have to push that pattern to its limit and that one person just directs more hours than exist right that's where I find myself and I found myself you know since kind of Claude Cowork and Claude Code kind of kicked off this phase in late 2025 early 2026 and I think Codex has really taken over the narrative ever since. But, you know, essentially I'm directing more hours that exist in a day, right? I'm literally getting done what used to take me even two years ago, right? Even in the early

chat TVT days, I'm getting about 30 hours of 2024 me work done every single day. So it's just, you're directing more hours, more traffic than can fit on a traditional runway. And, you know, So an OpenAI report came out with this. They showed how their team was using Codex. And what I found interesting was OpenAI staff, the 99th percentile, generated over 60 hours of agent turns every single day. So yeah, that's kind of the anomaly, right? That's not the average because they're building this technology. But I think eventually that is going to become the default. And that is going to force people, I think, to question, how do you work, right? What is the value of my actual time? And how can we responsibly, you know, use these agents as they get even smarter and even faster? I actually think we're in a nice little groove where we are today, right? There's been all this, you know, talk in the last, you know, 36 hours or so of this concept of AI pacing.

You know, essentially a lot of researchers, particularly at OpenAI, Anthropic, but also Google, Meta, and other labs, you know, kind of sign this letter saying, hey, we support, you know, essentially calling it like pacing, right? But it's kind of just means like, hey, it's okay if we slow down. Will that happen? Probably not, right? Is it the responsible thing for, you know, some sort of AI pacing to happen? Probably, because the issue is the capability gap is growing. The model overhang is growing, and it's not going to get any slower, right? Especially as, you know, over the past couple of months, we've gotten, you know, real results from the big labs talking about RSI, recursive self-improvement, where the models are improving themselves, right? Yesterday, OpenAI literally just announced like an 18% improvement via RSI. They essentially said, hey, we got, you know, we use GP5.6 to make itself better, and it's like 18% better. It's going to see, you know, 18% cost savings now. So this is where it's at. These models are going to get better and better,

faster and faster, but way at a much faster accelerated rate than humans can keep up with. So yeah, there's this whole pacing thing that maybe we'll tackle at a different time. But the brutal reality is if you want your team to be able to keep up and get ahead, that's something that you have to tackle right now head on. Some other good stats that recently came out, a Workday Foundation study in May found that 86% of users felt more productive. but those same workers describe the conversations that rarely moved past transactional work needs, right? And that's kind of where I feel right now. It's like, yeah, I'm getting way more done. The content I hope is good. The quality is good, right? What I produce is much different than what most people produce. But it's just kind of transactional, right? I've talked many times on the show about I'm not even because my, you know, the throughput, the output is so insanely high for what I'm doing on a daily basis as an individual full-time human, I don't retain information like I used to because I think so much information now is transactional.

And maybe that's something that we, you know, the future workforce needs to understand because I think, you know, have you ever thought about how you can learn a concept better if you write it down by hand or if you teach it to someone, right? I'm a big believer in you haven't really learned something until you teach, right? That's one of the reasons why I started a daily podcast. I'm like, I really want to learn this AI thing. I need to learn it well enough that I can teach at least one person, right? Luckily, hopefully it's helped more than one person. But I think the same is true. Because AI, it does all that hard work for you, right? You no longer have to go onto 20 webpages and read it all and be like, ah, this doesn't make sense. This does make sense. I'm not sure about this. Let me look at a little more. Now you just put something into whatever AI system you're using, you get a personalized, customized output. And well, most people don't even take the time to truly read it and understand it. So, you know, there is this disconnection between, you know, the productivity and disconnection that's shared in one workday.

And this study shows that that is the mechanism And I actually you know touched on some of these other concepts So you know if you do want to you know dive in a little bit deeper on you know some of these downsides of using AI the right way, I actually had an episode, episode 757. That was part of the Start Here series. It was called The Seven Silent Sins of Doing AI Right, How to Spot and Overcome the Invisible AI Work Traps. And three that I wanted to share that are kind of relevant to what we're talking about today is number one, accidental de-skilling, which is kind of one of the things I just referenced there. That's where AI does the work. So you lose the skill. I talked about the agent bun sandwich, and that's when agents start to hollow out your core expertise. And then the compression tax, where the AI speed overloads your brain. And that's where I think we might be getting into, right? With this, you know, this cerebras, right? Which might sound like a dorky niche thing. And You're like, okay, why do you keep talking about it? Because again, think, if you are an AI power user now, and I said, hey, tomorrow, you or

your company, your team can accomplish 20x what you accomplished today. If you are an agent-pilled team, and I'm like, you can accomplish 20 times more today, that's some compression tax happening right there. All right, so yeah, make sure to go check out episode 757. So let's be fair to AI, though. Some of this stuff, we don't need it, right? How many times have you gone through a three-hour meeting and you're like, that could have been a single bullet point? That could have been a single Slack message, right? Or waiting two days for a basic answer is not, you know, culture or mentorship. It's just maybe the signs of an antiquated business hierarchy that shouldn't have existed in the first place, right? I think most companies didn't quickly redesign how they worked for an AI native workplace. So, you know, you get to then decide which collaboration is intentional. You know, human to human, you get to choose, kind of pick your spots where that happens if you are AI native.

Okay, so as we wrap up here, I'm going to try to answer this question. How do you start to balance then of, you know, the reality of we're going to be managing more agents, which might lead to less human connection? And probably an increased expectation of productivity. How do all of those three things build together? And I know this isn't for 100% of organizations, right? I understand that. But I do think that more organizations, more teams will be shifting toward this than the opposite. So how do you meet that head on? Well, I think that winning companies are going to intentionally automate the manual work, but keep and prioritize human judgment. So I think generic human loop will always fail. So you have to name one expert owner. I always talk about expert driven loops and expert driven loops actually require instead of one passive human generalist, it requires active expert human collaboration and communication. So I think that's one thing right there. But also I think winning teams know which friction to remove and which human handoffs

to protect and maybe elevate. So then how do we start the balance then? Well, I think, yeah, as my slides are losing my place here. Sorry. Yeah, all screwed up. There we go. My slides got all out of order. So here's the three human handoffs that I think you need to protect. So number one, you need to protect the judgment by naming the human who owns consequences when the answer is wrong. Okay. OK, number two, you need to protect learning by making people explain decisions instead of blindly approving agent output. So that's kind of the difference between, like I said, passive, lazy, human in the loop, which will always fail and the correct way to do it, which is expert driven loops, which is experts communicating with each other, but also expecting inspecting agent outputs. And then last, you need to protect your belonging. Right. I think so much of this, so much of what we do in our work, it's about having this sense of belonging.

It's about, you know, some people, for better or worse, wrap their self in their own meaning is who they are in their job. Right. But I think you can still protect your belonging by giving coaching recognition in unstructured conversation, a scheduled place to happen. All right. So what's my big takeaway here as we get faster agents that are smarter and maybe we're just managing dozens or hundreds of agents and talking to our human co-workers less but producing more? Well, I think three big takeaways for me. Number one, everyone needs to increase their ambition. Right. Not just saying more prompts, more outputs, but being more intentional with what you can accomplish. right? Maybe not just doing 20x of what you were doing yesterday because you can do 20x today. Maybe instead it's thinking higher. It's thinking harder. It's just being more ambitious. Number two, you always have to unmarry your domain expertise. That is the, you know, the building blocks of what

I've been talking about now for three and a half years, unlearning. You have to unmarry your domain expertise because it doesn't really matter necessarily anymore. Your domain expertise is really just about that, you know, the human agent bun sandwich, right? It's making sure that you give it the right context and check it in the right way, but you're not going to be doing the 95% in the middle anymore. So you have to unmarry your domain expertise as agents get faster and smarter. And then last but not least, you have to be prepared to rebuild monthly instead of yearly. I think in the pre-agentic phase of 2024 or early 2025, you know, I was still recommending that people revisit their AI strategy roadmaps quarterly. But I would say even if you did it yearly, you know, in 2024, you weren't going to be left behind. Now you will. You have to be prepared to literally unlearn and rebuild how your company operates every single month, which sounds like an absolutely asinine concept because before you only had

to do it every five years, right? You know, or maybe every 10 years. And then, you know, as the Internet and, you know, Web2 and social media and all these other things started to gain in popularity and how people make decisions and all that, you know, advertising, marketing, communications, you know, those things changes. And so, you know, with digital transformation, I think that shortened that lifespan of, you know, how your company works, kind of your your company's IP, not just your SOPs, but literally how your company works and runs and thinks and makes decisions. you can't do it anymore. As crazy as it sounds, I think probably a majority of the time for successful companies moving forward are going to spend as much time unlearning and rebuilding processes as they are actually doing work the old fashioned way. All right. So I hope this episode was helpful. It's just more of a kind of a zoomed out look of saying, I don't necessarily know what the heck is going on or what the heck is happening next. But if you look at the writing

on the wall, you have to see and you have to understand agents are getting smarter. Agents are getting faster. And we, as people who are AI leaders in our respective organizations, well, we're going to be working with way more agents and probably fewer humans in many cases. So I hope this episode was helpful. If so, let me know. Please subscribe to the podcast. If you want to hear more or less of shows like this, do let me know by signing up for the newsletter at your everydayai.com. So thank you for tuning in. Hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all.