How I work with AI agents
I run AI agents on my client projects, often several at once. The testing and security checks that usually get cut when a deadline gets close now get done. Here's how I split the work.
What I never hand off
- Talking to the client. Every message a client gets comes from me.
- Deciding what to build now and what to leave for later.
- Anything that moves real money, like the first live payments.
- Going live. I do it one step at a time and check each step before the next.
What I check myself
- Every plan, before any work starts. An agent can draft it, and I decide what gets built.
- Every change, before it goes live. I review it and try it myself.
- Anything that touches payments, customer data or logins gets a separate security review before launch.
What I hand to agents
- Reading the request and the code, then coming back with a draft plan and a list of open questions for me.
- Building a piece of work after I've described it and signed off on the plan.
- Writing and running the tests.
- Checking each change against what the client asked for, and clicking through it in a real browser, before it comes to me.
Agents inside a product
An agent I build into a client's product works the same way. It has clear limits on what it can do alone, a person approves everything else, every step goes into a log, and there's an off switch.
Work with me
Client work runs through CodeFavorite, the practice I founded.
You'll find the work and how to start a project there.
Visit CodeFavoriteOr email me at hey@codefavorite.com.
