How I ship a SaaS with AI agents without writing the code myself

Why you no longer have to write code to ship real software

AI coding agents can now write, test and change real code. That does not remove the need for judgement. It moves the work. I spend no time typing code and most of my time deciding what to build, writing it down clearly and checking the result.

I built NIUJU, our AI portrait service, this way: one founder, a small set of AI coding agents and a fixed routine for checking their work. The NIUJU case study shows what came out of it.

Setting up an AI agent team with clear roles

One agent that does everything gets confused. A small team with clear roles works better: one agent plans, others build, and a separate agent reviews. The agent that reviews a change is never the one that wrote it.

Each role gets a short written brief: what it owns, what it must not touch and how it reports back. This keeps changes small and easy to follow.

Spec templates that make agents reliable

Agents do what you write, not what you mean. A good spec states the goal, the files involved, what must not change and how to prove it is done.

I use the same template for every task: the goal, the files in scope, what must not change, and the evidence that proves it is done. The fixed structure makes gaps visible before any code is written.

The verification workflow: build, run, confirm

An agent saying "done" is not proof. A change only counts as done when it has been built, run and checked against the spec, with evidence such as test output or a screenshot.

Changes that touch personal data, logins or payments get a second review before they go live. It is the same rule we follow when we build for clients, for example in a Secure Launch project.

Project instruction files that tie it together

Agents start every session without memory. A set of instruction files in the project tells them the architecture, the rules and the decisions already taken. Without these files, every session starts from zero.

What still needs a human in the loop

Agents are fast, but they do not carry responsibility. Product decisions, security and privacy trade-offs, anything legal and the final release stay with a person.

My job is not writing code. It is setting direction, asking the right questions and saying no. Seven years of reviewing whether other people's security controls really work turned out to be good training for reviewing an agent's work.