AI Agents are briefly overhyped

May 2, 2026

My dad asked me about AI agents:

I keep hearing about all these productive AI agents out there. We currently are prevented for security reasons, but it seems like we can be so much more productive using AI agents and there must be a way to do it safely.

– Rodger Krouse (edited for clarity)

Here's what I replied:

Right now, I think AI agents are overhyped outside of software engineering. This is especially true in heavily-regulated industries where security is important, and you can't give full-computer-access to the agent. But we should check back in in 3 months.

First, to demystify the word "agent", I want to start with a definition borrowed from my friend Simon Willison: "An LLM agent runs tools in a loop to achieve a goal."

I have this image in mind when I think about agents:

Source: Solomon Hykes, cited by Simon Willison

You're already using AI Agents

By this definition, every time you use any modern AI (ChatGPT.com, claude.ai, etc), you are using an AI agent!

At first, talking to ChatGPT was just Human <-> LLM, but then they started adding tools that the LLM itself could decide to call, in a loop.

The first tool that every LLM platform got was "web search", the ability to look things up before answering you. This was quickly followed by "code interpreter", which allows chatgpt or claude (reviews by Simon Willison) to write & run code on your behalf to do extremely complex analysis. So if you want to use agents, just pop open chatgpt.com or claude.ai (or their desktop apps) and give them very complex analytical problems, and they'll use dozens of tools to help you achieve your goal.

Claude Code, OpenClaw, Mac Minis

So what's the hype around Claude Code, OpenClaw, and Mac Minis?

Claude Code is an AI agent that can use almost any tool on your physical computer, which makes it very powerful, but also dangerous. If you think about it, if you're an information worker, most of what you do is at your computer, so it's an AI that can do almost anything you can do, from a work-output-perspective.

Claude Code is about a year old and reached $1B revenue runrate faster than almost any product in history. It is incredibly popular in developer communities.

OpenClaw is a wrapper around Claude Code that lets you run it 24-hours a day, so you can treat it like a human employee, texting it work, or having it do work while you sleep. People are buying mac minis on which to run OpenClaw, and thinking about it like buying a computer for a new employee. That's why there's a shortage of mac minis now.

But this is also why mac mini's especially don't make sense for your firm: you guys don't buy macs for new employees. You're a Windows shop.

There are also cloud-hosted versions of OpenClaw, which means you don't have to buy the physical computer; you can rent it. My friends at Kilo and Every both offer these hosted OpenClaw products, and many more do too. However, my advice would be for you to wait for these features to come in a more business-friendly format. Right now, they're too bleeding-edge. (The bleeding edge is where bleed.) It seems like OpenClaw hype has peaked, and folks are coming down to earth about their ability to productively manage agents.

(Note to reader that I didn't include to my dad: If you're technically inclined, I agree with Fly.io that You Should Write An Agent. My version is ~30 lines.)

Everyone will become a manager of their old job

AI agents will continue to get smarter and more powerful. In my company, this means it no longer makes sense to hire interns or new grads as software engineers. It's easier and cheaper and faster for us to manage AI agents. That's pretty crazy! At the same time, the value of super-senior software engineers is higher than ever. Big tech pay top talent millions and millions in salary to programmers. We're still hiring, but only the best of the best, like everyone else, so there are very few of these super-smart-folks to go around. The market for intelligence (human and machine) is going to get wonky!

I think this sort of thing will happen in every industry over the next 2 years. Everyone will realize that they can delegate more and more of their work to AIs, and become managers of whatever job they used to do.

Does this mean we can do more work with fewer people? In some cases yes, and some cases no. AI might make every company more productive, which means from a competitive perspective, nobody is ahead except the customer. Currently, there is a very small limit to how many AI agents 1 person can productively manage. I can only manage 1-3 personally, and not all the time. Anyone who claims they can manage more is fooling themselves, I think.

My recommendation to you is to empower and encourage employees to act like managers of the job they used to have. Give to AI the tasks that they are given, and manage the AI. See where it helps and where it doesn't. And keep trying, because that'll change every 3 months. I think you're basically already doing this, which is great!

FAQs

In reading this as a blog post, I think folks might have questions:

1. What about junior people? How are they supposed to find work? How can we expect to find more super-talented folks if we don't invest in training the next generation?

It's hard. Being able to write decent code no longer gets you a 6-figure-job.

This is sad for CS new grads but really good for the world! Now that power is democratized to everyone for so much cheaper!

But the question still remains: how should new grads get jobs?

I don't know, but I'd love to try to help. If you are a new grad or someone struggling to get a job, but aren't afraid of working hard, email me at steveykrouse@gmail.com. I'd be happy to give you some advice. I am very easily teacher-sniped.

2. What if I don't want to be a manager of my job? What if I like doing my job?

A lot of us software engineers are grappling with this question, because many of us love the act of programming, and see ourselves as craftspeople. I've written a couple posts on this topic, most of which mostly hold up, despite improvements in AI which mean that I too find myself doing more management and less coding:

Ultimately, it is my hope that agents are just a stepping stone towards better tools, so we all can be craftspeople.