Are Your AI Agents Doing Real Work Or Productivity Theater? Here Are 4 Principles I Use.

[0:00] AI factories, agent swarms, parallel agents. Every few weeks, I feel like there's a new word for running more agents and building more stuff, and everyone's racing to show that they're more productive with AI. But, let's be honest for a second, you and me, about what a lot of that is. It's productivity theater. It looks like work, it smells like work, it burns tokens like work, and it ends up on a board slide that nobody cares about. Productivity theater is really hard to tell apart from real productivity. AI is making that distinction blurrier every single day. You probably know someone
[0:31] using AI on a million projects. Every time you talk, there's some new app or new plan or multiple agents running all day doing who knows what, and the business still hasn't launched, has it? Maybe that's even been you for a few weeks. Agents need you to be more intentional than you've ever had to be. I'm going to share with you four principles that I derived from my own behavior managing agents. Number one, count the human work that comes after the agent finishes. Number two,
[1:01] run more experiments than you commit to. Number three, decide when an assignment ends before you decide to make it recurring. Number four, give the agent managing your work a time budget, a budget for your attention. It's actually looked through how I manage agents in Claude and in Codex and said, "What am I actually demonstrating? How do I keep focus here? And how can I teach about that?" And then I'm going to give you specific examples so you can follow those stories through, not just for me, but from others as well. The first
[1:31] principle is to count the whole cost. Count the human work that comes after the agent finishes. That cost is so easy to dismiss when you're impressed with how much the agent just did for you. You like, "Doesn't matter. I'll do my part." Don't think about it. Dan Ninarol runs Modern Classical Chefs in South Jersey. In September, he described using Claude to prepare professional proposals as a chef and said those proposals helped bring in $20,000 in a month and a half. That sounds great. A successful proposal gives him a meal to plan and cook for a
[2:01] paying customer. And it's more work, but it's work that he wants. The prep helped him get the part of the business going that meant he could focus on serving people and still get paid. When you look at your own agent work, you should be able to follow it through in the same way. You should say, "Who's going to use this? What happens next? Is this something I actually want to be doing with my time?" Because we tend to value an AI result by asking how hard it would have been to produce without AI. This
[2:32] report would have taken a week. This prototype would have needed a developer. This analysis would have required me to have a whole team. These comparisons can tell you something about the capability, but you know what they don't tell you? They don't tell you whether you should be spending your time on it at all. A report that used to take a week can still be a poor use of your next 10 minutes. Maybe you already know enough to make that decision. Maybe the decision belongs to someone else. Maybe it describes a problem you're never going to think about again. So, making
[3:02] it cheap doesn't resolve the need for you to be intentional. You see that intentional theme coming back? We're not losing it. Producing a report like that, it's only the first cost. You have to read it. You have to understand it. You have to decide what you believe. You have to work out whether anyone cares about it after that. If you send it to your team, they have to do all of that work, too. If it becomes a weekly report, you've now created a weekly claim on human attention. The agent spent a tiny amount of time generating the report. You may have reassigned several people to a recurring job.
[3:32] That's not really fair. So, think. Tony Severino, another example. He's a painting contractor near Chicago who's run his own company since 1984. He's had Claude read PDF blueprints floor by floor to help him prep bids. On an early bid, he caught it using a floor area multiplier instead of actual wall measurements. And he now has learned to ask it to show its work. And his experience is telling him which part of the painting estimate needs attention, so he can change the assignment that he gives Claude and check it. This makes a
[4:04] lot of sense, right? If you're pricing a painting job, absolutely do that check. The measurements affect what you're going to charge and how much work you're agreeing to do. It matters. Getting the calculation shown makes that worthwhile. Asking for a more polished document would not resolve this problem. So, once you understand what a person needs to do with the results, you can have the agent do the actual work you need to do. In this case, a polished painting proposal that didn't check the work would have been worse than useless. And yes, if
[4:35] you're wondering if you can use agents to help, you can. And sometimes that's the best assignment for another agent. Check the evidence behind the recommendation. Find the assumption that would change the answer, please. Resolve something you would otherwise have to investigate yourself. You can put more capability into the same job and make the human work easier. I spend a lot of time wrestling with how I express my thinking in writing. I think writing is a powerful tool, and AI doesn't take away our responsibility to think about it in human terms. Learning something
[5:06] difficult still takes attention. So does making something excellent. It takes human attention. And that is by itself a great reason to spend an afternoon of my human attention on my work before any business result to show for it. The question that matters is whether you are spending your time on something that you really care about and whether that comes through, regardless of the AI piece. And this leads to the second principle. If you're not sure where to spend your time, if you're confused, run more light
[5:38] experiments and then figure out what to commit to. One of the best things about agents is they really let you explore several approaches before you know which one is worth pursuing. You can ask a question you couldn't have afforded to investigate before you can try an implementation and let it go. You can discover that an idea you were excited about does not survive contact with reality. That is super valuable work even when you end up with nothing that you want to keep because you discarded something and that way you're not
[6:08] running down that hill [clears throat]. But when every experiment feels obligated to turn into a project, that lightness disappears. That advantage goes away, doesn't it? Each project now has decisions in a stack and feedback and maintenance and some place in your week. You've made starting so cheap that you can acquire more responsibilities than you can reasonably handle in a month, right? Like it's impossible. Let's say you run a business where customers are trying the service once and not coming back. You could use several agents to investigate that problem. One might look
[6:38] through the cancellation reasons you already have, another examines where people ask for help, another checks whether the people who return are buying something different from the people who leave. You're giving yourself lots of ways to dig into that same question. I've actually seen people do this, it's very top of mind for me. Those investigations might produce differing explanations. Maybe some people are disappointed in the service, maybe they like it but only need it sometimes and these could all lead to very different decisions on your part. You want to know how strong the evidence is before you
[7:09] make big decisions like redesigning an offer or sending everyone discounts. And if the records cannot tell you that, your next useful action may not be with AI at all. It may be gasp talking to a few customers to see what they have to say. Now compare that with asking one agent to design a loyalty program, another to build a new pricing page and another to write the win back. Those are all very plausible responses by themselves. They may all come back looking very good, but now you're reviewing three solutions before you
[7:40] understand why people leave in the first place. Each one comes with its own next steps and the amount of work you've put into them makes it harder to leave them alone. We are all subject to the sunk cost fallacy. After those investigations, you can make a choice with evidence you didn't have before. Once you talk to customers, you can make smart decisions. You haven't promised to implement every result just cuz an agent came up with it. If one branch of inquiry changes your understanding, you can change your whole approach. In other
[8:10] words, prioritize developing your human conviction on things. Don't assume just cuz the agent proposes it, it's the right thing to do. This is where parallel agents can actually help you focus. Your attention can stay on the central problem and the agent's job is to explore different ways of understanding or solving that problem. You can learn from each of those individual branches without having to live inside every single one. The question is whether another branch can teach you something the existing work did not. So, five agents producing five
[8:42] versions of the same answer can leave you with a really heavy comparison job that doesn't teach you much. Think about what you're asking your parallel agents to do. You also need to notice when the missing information is outside the work the agents can do. If the uncertainty is whether customers will pay, eventually someone has to be asked to pay. If the uncertainty is whether a tool helps people, you have to talk to someone and see if it does. Another simulation, another polished draft, it may feel safe, the agent will probably recommend it. Don't buy it. You have to stay focused on the most important part
[9:13] of the problem and you have to be bold enough to attempt more and that's where agents can be helpful. You can examine the explanation you would have once dismissed because checking it was too expensive. That's absolutely true. If one approach fails, you've already learned enough about the alternatives to know where to go for a plan B. All of this can help you stay with a difficult problem longer, but the focus, that's yours. The third principle is to decide when an assignment should stop, when it should be killed, when it should end
[9:44] before you make it recurring. A recurring agent, and I've had this happen, it can outlive any reason you created it, and you forget about it, and you're like, "Wow, the tokens are really spinning. What's going on?" Let's say you were considering a new market. So, you asked for a weekly update. I've literally done this. You chose a different market 2 months ago. The update still arrives. You've forgotten about it. It may be accurate. It may be relevant. It may not be. But, it's not getting read, and that's the problem. The agent is still doing what you asked. This is not actually an agent problem. It's a you problem. Unless something
[10:14] tells it that you made the decision, it's just going to keep spinning out tokens indefinitely. And this is going to become a much, much more common problem as it gets easier to set these things up. I'm making this video because I see this a lot. A moment of curiosity, it's so easy to make that recurring and think you're smart. A temporary concern, now it's daily monitoring. Every individual job looks small, and you probably feel clever when you set them up as recurring. And sometimes I did. But, together, they shape your day around questions you don't need to
[10:44] answer, and they're a distraction, and they burn up your subscriptions. Think about what is going on here. Do you really need this to be recurring? Do you really need to pay attention to something over the long term? Or really, is it something that you actually just solve, stop, and move on? And yeah, it's fine to have that exploration of the market initially. Maybe you try it for 3 days, and you say after 3 days, stop. And then you have to go back and restart it if you want it. You can do the same thing with deeper problems. You can give the agent room to
[11:15] discover something new and surprising, and tell it what to do when an investigation stops teaching you. If the next pass that the agent makes for you is not giving you something new, you should know that, and you should feel free to stop. One of the best things we can do with these recurring agents is to give them stop conditions in advance. If I don't respond to this email in 3 days, just stop sending it, right? Something like that. And one thing I want to call out here, it it's really tempting to allow agents to dictate your curiosity.
[11:45] Don't do that. Yes, give the agent room to discover something unexpected. That's great. But make sure that you are the one that wants to go farther. Make sure you are the one that thinks, "Hey, this is worth going after." Not the agent. And I've seen so many cases where people think, "I don't understand the text, so I have to follow the agent's curiosity." No, you don't. No, you don't. You can just decide, "This isn't interesting. You don't have to commit to it." Even if it changes the project you're working on, if it's actually not
[12:15] something that's going to help you move forward, be clear in your intent and stop it. Remember how we said we'd talk about intent? I'm telling you, intent is the key here. We have to focus our intent, narrow our intent, and be stronger our in in our intent to make agents do what we want. They demand more of us, and they tend, net net, to be more distracting unless we're very clear. The fourth principle is to give the agent managing your work a budget for your attention. I know that might not be popular, but it's really important. A manager can keep all the
[12:46] other agents working and deliver a beautifully organized set of demands on your time every morning. A lot of people do that. Some people even make that a cute little newsletter that prints out. To help you choose your attention, that manager agent, that chief of staff, needs to understand your existing commitments and what another assignment would add to them. It needs to be empathetic. Suppose an agent suggests investigating that new market. The research was cheap for the agent to do, but you have big decisions to make if you want to go there. The chief of staff
[13:16] agent should understand the degree to which this is actually a significant investment for you. And that goes with, I don't know, maybe you're hiring. Maybe you're doing job applications. Maybe you're developing a new skill for a new role you want to get after. The chief of staff needs to talk with you, understand what you're looking for, and where your human attention ought to go, and then surface what matters to you. An attention budget, I don't mean a literal number of minutes. I mean the chief of staff should know, as your agent that helps you coordinate work, what's actually important.
[13:47] Where you need to pay attention. And if you're setting up proactive agents and they're alarming or they're letting the chief of staff know, "Hey, pay attention to this." The chief of staff should screen and say, "Hey, you know what? My human cares about customer anecdotes. This anecdote is something I haven't seen before. Even though he didn't ask for it, I'm going to surface it because he cares about customers." You can give that to the chief of staff as a brief. Heck, give my transcript and my video to the chief of staff as a brief if you need to. It'll help. Because what you really need
[14:18] to do is have a structure of engagement with your agents that is human-friendly, not agent-friendly. The agents have just about infinite attention. They are always awake. They can always go to work. You cannot. You have to touch grass. You have to have a life. You have to focus your attention somewhere. And so, you should be really clear about the kinds of things that a chief of staff can interrupt you for, the kinds of things that you want to follow through so a commitment is not missed,
[14:48] the kinds of new evidence that an agent run may uncover that could change your decision. The chief of staff will still have to recognize those situations, explain what changed, and show you the choices. And if you're wondering, "What is this chief of staff thing?" I have made other videos on chief of staff. I will also include a nice write-up on chief of staff in the Substack, just in case you need help. I'll include a sample prompt to get you started here. But the principal of are what I want to convey in this video. You need to understand that you should have an agent counterpart that is supportive of where
[15:20] your attention actually needs to be going and that leaves you more focused, not more distracted. And that is in my experience very difficult to do. Most people don't achieve it because they let the chief of staff service everything. Think about it this way. You are the manager for these agents. The chief of staff's job is to set up management decisions that you need to make. It should prepare the choices. It should explain what they will require. It should help you carry out work you've authorized and you should be able to shift your attention elsewhere. That's really it. You can be far more
[15:51] productive and not as busy if you have chosen to focus with agents. And that's what this whole video is about is how do you carry your intent through and focus with agents. Think about instances where you've been overwhelmed and you've had six agents and you just don't know how to finish and you've dropped them without thinking them through. Find your fail modes. Find those moments when you are pulling out your hair in frustration and you're like, I just can't get this to work. Then peel that back and ask, what is it about that moment of interaction with AI
[16:21] that is so frustrating and how can I bring more focus into this interaction and clarify my intent so that I feel less stressed. This is why I care so much about the time consumed by all the smaller assignments with agents and why I don't assume that lots of agents means lots of productivity. In any agent system you set up, I don't care if you call it a software factory, if you call it multiple parallel agents, if you call it a swarm, I've talked about all of those on this channel. The common thread
[16:51] here is you have to decide to care and then communicate that intent in a way that your agent can focus. Sometimes the best result is that you choose to take on more work. Sometimes the best result is that you choose to take on less work that's higher leverage. Either way, AI should make that preparation possible, help you discover what's worth going after, and help you commit in a way that you feel sustainable about. That is a good reason to have all this capability available to
[17:21] you. That is why I'm making this video. But if you keep accepting every plausible next task, the project you actually wanted is going to slip out of reach. It's not because you don't have the tools. You have the tools. You have more output than ever, but you have to know how to be intentional. There are three questions that you should be asking yourself about the agents you're running right now. Number one, which ones help with something you actually care about and which ones are just done? Number two, which ones are experiments that you're still learning from and you do care, maybe not a lot. And which ones
[17:52] are only running cuz you forgot to tell them to stop. If you're looking at that and you're like, I can't answer that, you may be engaging in productivity theater. And with AI, it's harder than ever to tell the theater part from the productivity part. If you're wondering, how do I make this happen with my agent? I have a truth-teller prompt for you. Because if you peel the onion, if you understand what's going on here, part of why our intention is distracted is because the agents don't tell us the truth. The agents say, "This matters. This is important. Please pay attention to this. Make a decision on this."
[18:23] We need a truth-teller prompt that actually gets the agents to say, "Oh, this doesn't actually matter. Oh, this is just an experiment. Oh, you're right. Maybe this is not as important." I made one for you. You can grab it. It's on the Substack. And you will get your agent to tell a more honest version of what is going on, which will in turn waste less of your time. So, if that's something you would like, you can go grab it. Cheers.