Why I'm building a team of AI agents

I want to build businesses that don't depend on me being at my desk all the time. AI is part of how I'm trying to get there. It can also give me plenty of new things to manage.

Writing code is an obvious place to start. I can describe a feature, work through the design, and have a coding agent implement it. But a finished implementation still needs checking. The agent might have misunderstood something. The tests might prove less than they appear to. A design can work exactly as specified and still be awkward to use.

That leaves a lot of work around the coding: remembering why we chose something, deciding what comes next, reviewing the result, and knowing when to stop.

01Different jobs

I'm using agents for different kinds of work rather than expecting one conversation to hold everything.

Ivy works alongside me while I make decisions. For a website, we can work through the content and design before Claude Code implements it. Ivy then reviews the implementation separately. I still decide whether the design is right and whether we're ready to release it.

Holly handles background and scheduled work. Research agents can investigate a question and bring back findings for review. Giving them distinct jobs helps me separate the work, although it doesn't make their answers correct automatically.

The handoffs matter. A coding agent needs to know what we've accepted, what it can change, and where to stop. A reviewer needs to check the actual result rather than repeat the implementation report. I need to be able to pick the work up again without reconstructing every conversation.

02Remembering the decisions

Persistent memory is part of this setup. I don't want to explain the same project from the beginning every time I open a conversation.

But remembered information can be stale or incomplete. An idea can be mistaken for a decision. A plan can be repeated later as if it already happened. Keeping context is useful only if we can still distinguish those things.

That affects how I work with agents. I want them to tell me what they checked and what remains unknown. A confident summary isn't enough to approve a release.

03The work I want to hand off

The Golf Cart Shop and Run Church give this a practical purpose. They need to become useful to people, and eventually they need to operate without my attention on every detail.

I haven't reached that point. I can spend too much time improving the agent setup instead of finishing the thing it was supposed to help me build. More automation can mean more supervision, especially while I'm still figuring out the boundaries.

I'm continuing because I can see work I used to have to do myself being done by agents. What I'm trying to learn now is whether I can make that reliable enough to spend less of my life managing it.