How I work with AI
This page explains how PT Max was built with AI coding agents, and how that work was kept safe.
I'm the architect and technical lead in the team behind PT Max, a live fitness-coaching platform delivered through Nhance Group. Much of it was built with a team of AI coding agents (Anthropic's Claude Code) working in parallel, and nothing reaches production without a person signing it off. You can see it at ptmax.app and on Google Play.
There have been 6,394 commits since the first one on 2 February 2026, and Claude wrote 5,205 of them (81%), each co-authored by the agent. Ten or more agents usually work at once, each with its own job, branch and working copy, and as many as 13 have worked on it in a single day.
The code is covered by 9,456 automated .NET tests and 1,381 browser tests. The full browser suite runs before every release, and it scales to use every core of a modern computer, with successful runs on up to 32 in parallel. PT Max has been on the Google Play production track since August 2026, it takes card payments through Stripe, and real coaches and their clients use it.
The harder part of working with AI agents is trusting them with a live product that people pay for. I designed the rules and safeguards that make that possible, and they fall into five areas.
One coordinating agent shares out what the others can't all use at once: ports, the development database, test phones and the production server. It also keeps a record of every decision. Each working agent owns its branch, merges its own work and reports back. Anything that can't be undone needs a person's direct approval, and an agent refuses such an instruction if another agent passed it on.
Five operating levels, from normal running to "production is down", change what the agents may do, and a code freeze is a separate lock on top. Deploys, anything that moves money, messages to real people, deletions and secrets always need a person's approval.
More than ten thousand automated tests check the code the agents write, and the full browser suite runs before every release. A contract test catches any change that would quietly break the version of the app already on people's phones.
An agent may change any test it needs to, but it has to list every change openly, with what the test checked before and after, why it changed, and what would make the change the wrong call. A report that an agent found nothing doesn't count as evidence until the agent shows its check could have found something.
An agent raises it with a person before any change that could lose real data, any bulk edit, and any notification a real person would receive. A development server can't send a real email, by design, and every production deploy is run by a person rather than an agent.
I built an operations dashboard for the agents and the product, which shows production and staging health, test history, certificates, errors and the release pipeline. A voice channel lets the agents report and ask for decisions out loud, on the speakers or on a handset, with priorities and quiet hours, so decisions don't have to wait for someone to be at a screen. The agents also gather their open questions into decision pages that are answered by clicking accept, reject or comment, and they read the answers back themselves. For private material that shouldn't go to a cloud service, I use a local language model on my own machine.
AI agents can make an engineering team considerably faster, and they can just as easily ship mistakes faster. I've learnt how to get the speed while keeping control of quality, security and production, and I can bring that to your team as a fractional CTO or a senior engineer.