Someone asked me how I manage working with more than one AI agent at a time. How do you keep track of what each one is doing?
I said I have to take a detour and tell you about waiting tables.
When I was 19 I got a job bussing tables at Stone Brewing’s first restaurant in Escondido. When I interviewed, they asked me what my favorite beer was. I said “I can’t drink, I’m 19.” They said “good answer.” Out of 60-ish beer-loving new hires, I was one of three who couldn’t legally drink, and the only one who didn’t want to.
A year and a half later, after turning 21, I started training as a server.
Each day, I drove in, got the new bar menu, headed home, and looked up every beer on BeerAdvocate. I made notes in the margins of the paper menu. If someone at my table asked for a nutty brown, I could point them to a few options out of the 50 or so rotating beers we had.
Then, during my shift, I’d take an order at one table, recommend something I’d researched. As I walked away, the next table flagged me down for a side of barbeque sauce. On the way to the kitchen to enter the first order, I’d see a third table getting set out of the corner of my eye, and I couldn’t remember if I was supposed to bring menus to a table that just added a person. I never had time to write any of it down.
I don’t remember a specific moment where it got easier. My tips started improving, and it didn’t feel like much had changed, but something had. I wasn’t stepping through the checklist anymore: where’s the barbeque sauce, did table four get menus, is the kitchen backed up. I was listening to what people were telling me, and seeing what I could do to make them happy.
Still, on Sunday brunches I’d get sat with too many tables at once and have to ask the host to stop seating me for a bit.
One instance
When I started using Claude Code, most of my energy went into managing a single session. “Can I read this file?” Yes. “Can I read this one?” Yes. “Can I read this other one?” Yes. “Can I post a comment as you on GitHub?” …hold on.
I bounced off completely the first time. I came back when it still asked for a lot, but infrequently enough that I could actually work.
Dead time
Once I set up my permissions settings, Claude started running longer without interrupting me. Every five or ten seconds became every minute or two. Two minutes feels like a long time when you’re watching a spinner.
Say I’m waiting for something to finish. Someone just reported a bug in Slack. So I open another Claude session. Here’s the Slack thread, and what I think is going on, look into this. I don’t need it to interrupt me. I just want it going, and I’ll check on it in a few minutes.
That was my second table.
Different cadences
Once I got the hang of this, I’d regularly work in parallel. One session is debugging. Another is researching: What’s the prior art for this design? What are people using for auth? The debugging thread might come back in thirty seconds. The research thread won’t need me for ten minutes.
I’d spin up sessions and forget to come back. A colleague pointed me to a page listing open pull requests. I had thirty-four open PRs at the time. I’d find myself saying “I have a PR for that.” How useful is an unmerged PR?
I was busy all the time and making little progress on the important things. Someone pulled me aside and said I might be taking on too much.
The next table
I took each incremental step because I got sick of something. Sick of approving every file read, repeating the same instructions, or copying Claude’s output into one window, reformatting it, and pasting it into another.
I still spend some of my time approving obvious things. I’m the middleman between Claude and Slack, between Claude and a design doc. That’s not a job for me, that’s a job for a computer.
So, coming back to the question that started this post: start with one. Then get a feel for the second before you pretend you can handle five.
I didn’t mention the part where someone had to tell me to stop.