Back to articles
engineering

Letting non-engineers ship to production with AI agents

By Daniel Chavez

I'm the technical lead for a franchise company with nine websites across four brands. This article explains how we set things up so people who are not engineers can change those sites with AI agents, and which parts I keep locked.

How it worked before

When I started, the designer built pages by prompting ChatGPT, exported a zip, and a developer wired it into the real site. So every change went through two people, and things got lost between them.

The owner asked for something simple: anyone on the team should be able to change the site directly, without a handoff.

I think that is the right goal. The problem is the lead form.

Why the form is the risky part

On these sites, a form submission goes to a database, then to the CRM, and then it starts an SMS follow-up. If someone renames a field in the page, the form still looks fine and still says "thank you". The lead just doesn't arrive, and nobody notices.

A wrong headline or a wrong photo is different. You see it on the preview link, and if it gets through you revert it.

So I split it in two. Anyone can change the front end. The code behind the form needs my review.

One repository and one path

We moved everything into one TypeScript monorepo. Shared components are in one package and each site is its own app. A new brand site starts from one command and gets the same form and tracking code as the others.

Every change goes the same way, no matter who makes it:

  1. Create a branch.
  2. Prompt the agent and check the result locally.
  3. Open a pull request. Vercel adds a preview link to it.
  4. I review it and merge.

The designer still uses a ChatGPT site to try ideas. Now an agent copies from it in a normal pull request, and nobody ports a zip by hand.

Rules the agent can read

An agent follows what it finds in the repository, so I wrote the rules there.

There is a set of agent skills in the repo that explain the project's conventions. There is also an instructions file that says who owns what. For example, the CRM belongs to the CRM specialist, and a commit by itself does not authorize a deployment or a live webhook test.

Without that, if you ask an agent to make the form shorter, it makes the form shorter and breaks the pipeline. With it, the agent tells you the form fields are a contract and points you to the file.

Checks in CI

I don't trust rules that are only written down, so the important ones are CI checks:

  • Form fields. CI compares each site's form with the intake contract. If a field is removed or renamed, the pull request fails.
  • Deployments. We had pushes that never deployed and nobody saw it. Now CI checks every push.
  • Unused dependencies. Agents add packages all the time. This check catches the ones nothing uses.
  • Tests. When I moved the SMS automation to a new host, coverage went from 75% to 83% across 623 tests.

API keys stay on the server, one set per environment.

I still read every pull request

The agents write code that looks right, so the bugs also look right.

The best example had nothing to do with an agent. Google was throwing away the client's ad conversions. The new sites were not sending the ad click ID with the lead, and the CRM was not saving it from any site. Every form worked and every lead arrived, so nothing looked broken. I only found it by reading the handler and following one lead from the form to the CRM.

How it went

Three sites went live in the first month. The staff now make their own changes, and the client doubled my hours within five weeks.

If you want to try this, start with the question I started with: which parts can be wrong for an hour without costing money? Open those to everyone, and put a check and a reviewer in front of the rest.