My AI team had senior skills and still missed the basics
For a while I’d been thinking that, in theory, I could run an engineering team made entirely of AI agents. A team lead, a front end engineer, a UX designer and QA. Tickets get built, go through QA, I check them visually, and they get released.
This week I was like, do you know what? I’m going to actually do it, and point it at my website.
So I set up the team and got my chief of staff agent to go find the best skills it could for each role and load them in.
The handoffs worked. I asked the team lead for a few fixes. It handed them to the front end engineer, who handed them to QA. I checked a preview, and it went to production. Honestly, that part was kind of amazing.
The judgement is where it fell apart.
Nobody thought about mobile. The page was falling off the screen on a phone. So I pointed that out, and then they swung completely the other way. The front end engineer was going to build for every iPhone model, and QA was going to check every single one. Complete overkill.
Then I asked the UX designer how long its rebuild would take. Nine to fifteen days of straight engineering build work. On the plan I pay for, that’s just never going to happen.
And the team lead started making the current site load faster. But I was rebuilding the entire site with the UX designer. Why would you speed up a website that’s about to be thrown away?
None of this is hard. Responsive layout, a sensible test scope, a budget, the order of work. It’s what I’d call basic logic. I’d asked for the most in depth skills for every role and still had to catch all four myself.
What I took from it: a skill tells an agent how to do a role. It doesn’t tell it what matters on this project, what it costs, or what else is going on at the same time. That’s the product manager’s job. The agents just didn’t have it unless I gave it to them.
So I changed the setup the same day.
Standing specialists were expensive. The UX designer alone used about a fifth of my weekly usage. Now there’s one standing team lead, and short lived helpers that only load a role skill when a job actually needs it.
Self review didn’t work. When the lead skipped design and QA, the result was poor. Now nothing gets to me without a fresh design pass and an independent QA pass, and QA only sees the brief, the design and the preview.
Constraints go in the brief: budget, devices, and what’s about to be thrown away. New work stops at 80% of my weekly usage, so there’s always room left.
I also restructured the team twice in one day, which is its own lesson. So the setup is frozen now, and I’ll judge it on output: better previews, fewer fix rounds, reasonable cost.
The part I’m still chewing on is this. If agents have less judgement of their own, I have to define more upfront: the scope, the devices, the budget, the sequencing. Writing everything down before work starts sounds a lot like waterfall. Are we heading back there?
I don’t think it’s that simple. What got heavier was the brief, the context. The loop itself stayed short: one small slice, a preview, my check, release. That still feels agile to me. But I’m watching it, because the more an agent depends on what I write down, the more a bad brief costs.
It’s not perfect. But a team with strict handoffs, short lived helpers and one lead reporting up to me does work, which is kind of super scary and interesting at the same time. The bit that’s still missing is the bit I’ve spent my career doing.