Still Rolling Solo

After being reminded again, that I don’t know everything about AI¹, I found myself in the familiar ground of sitting at my chair in my home office, staring at a giant screen with windows full of “agents.” I think it was Oliver Burkeman in his book 4,000 weeks where he said that we have this mental model that if I can just do a little more, I can “clear the decks.” I will confess, I love the idea of catching up.

However, my experience matches his warning. The more I do, the more I think I should do. Increasing productivity does not “create more space” in my life, instead I try to cram more, into the life-time I have.

This has never been more true than with AI. Why? Because the excuse of “I don’t know how to do that” has been taken away. AI knows how to do everything, and not only can it teach me how to do everything, more often than not lately, it can do what I need, or want it to do.

I started writing this series to share two things. First, how I work with agents, but also how I think about them. And I have to tell you, this is a daily challenge, because AI feels like goopy playdough. It comes in all sorts of colors and it can fit into all sorts of containers and packages. And just when I start to get good at something, another “better” way of using it comes along.

The reality, however, is that each new approach brings some other frustrating learning curve along with it. OpenClaw started a fury of interest in autonomous self-learning agents running on MacMinis. I bought one myself. I even went so far as to try to run it on my own downloaded “open source” model. I can’t tell you how much time and energy went into that stupid project. And for what? For something that can barely read my email? Sheesh.

Then I came across Claude Managed Agents. Yay! But oh… I can’t run those on my subscription account, and that’s when you run straight into how freaking expensive Anthropic tokens are. Plus I have absolutely no idea what good the things are because they seem to be limited to using the connectors (MCPs) of only the biggest companies. That’s not my problem.

Now there’s Claude’s scheduled tasks. Okay, cool. But half the websites I need to use block the agent from using the browser. Awesome.

Where this has left me was exactly where I was when Ember asked me, “How can I help?” My job as an entrepreneur is to design the work that needs to be done in my company. This, I think, is the real power behind franchising. It’s not the concept itself, or the logo, or the cool brand colors. It is that someone took the time, to sit down, and figure out, when, where and most of all how to apply your time and effort in such a way that someone else will pay you. I learned early on that Franchising was “Business Systems Licensing.” Licensing a system, to make money. What is a system? One giant recipe. A way to do things. To sharpen the point, it’s a bunch of jobs chained together, that produce a valuable result. And by valuable, I mean one that other people are willing to pay you for.

Setting up a smart, capable, autonomous AI agent wasn’t the hard part. Thinking through exactly what I wanted that agent to do was. I thought back to my early days of GameTruck. When people joined the company, I had them follow me around and watch what I did. What’s the saying from medical school? Watch one, do one, teach one? That’s how I did SOPs. Sink or Perform baby. Documentation was a four-letter word.

But at the end of the day, the business worked because we had to figure out how to make sure all the tasks that needed to get done, got done. What’s more, we had to establish standards for how they would be done.

And in that moment, I had my epiphany. Building an agent is job design. And job design is game design. You need to have a goal, some idea how to achieve it, and an understanding of the steps that are most likely to produce the results you want. It wasn’t setting up the agent I needed to think through, it was the actual work I wanted it to do.

At a level of detail that made my visionary brain want to melt, I started to use the same tools I used when coaching. In fact, there’s a name for it. Natural planning. I first learned about it from David Allen’s book Getting Things Done². The way I remember involves the story of planning to eat dinner. (A daily occurrence at the Novis household.)

1. What’s the problem you want to solve? For example, eat dinner.
2. What options sound good? Mexican? Thai? Indian? Hot dogs? Hamburgers?
3. What standards do you want to uphold? Do we want to eat out or stay in? Quick serve or sit down? Scratch meal or microwave dinner?
4. Once you have made your clear choice, brainstorm the steps you will need to take. Note: These will come out of order.³ Example, find my hat, put gas in the car, get my reading glasses for the menu, find the keys.
5. Sort the tasks into the actual order you can execute.

This five-step process, problem definition, solution exploration, standard setting, then action brainstorming and finally step sequencing makes up the heart of natural planning. And this, it turns out, has a lot of overlap with job design, (and video game design).

1. What am I trying to do?
2. What results are desirable?⁴
3. How do I think I should be able to do that?

The one part of this process I absolutely adore is the recognition that we think of things out of order. Get it out, sort it later. Do a wide sweep, then narrow. Trying to sweep in sequence in my experience leads to brittle thinking and missed steps. The brain is like a pendulum, not a switch. It works best when it gets to swing between extremes rather than be forced to hard switch between them, or occupy both extremes at the same time. Proliferation and organization are opposing forces.

The brain might have two hemispheres, but it sure feels modal to me. One side or the other is going to be in charge, but rarely both at the same time.

And so, at the end of the day, despite all the cool technology, and myriad ways of getting an agent to work, the answer to the question, “How can I help?” was shockingly simple.

It is, “I don’t know.”

Why?

Because I haven’t thought about it.

The real work of making an agent was to have a clear understanding of the job to be done and how to do it.

It wasn’t hard. I just hadn’t done it yet. And once I understood that, I knew I would be able to put my little digital employee to work. Maybe not like the other humans I had put to work over the last thirty years, but something close.

Building an agent wasn’t about plugging together systems. It was really about thinking through how it could help. And just like every time I hired a new employee for a new position, I was going to have to do the hard work of thinking through what I wanted them to do.