How I Built Agent (3): From Chat to Codex: When the Way You Work Changes

At first, I developed Agent in a rather old-fashioned way. ChatGPT suggested a file, I saved it, uploaded it to the server and waited to see what happened. It felt like a game in which I knew the rules but not the location of the next obstacle. As the number of files and dependencies grew, this approach became difficult to manage.

From chat to repository

Moving to Codex and GitHub changed much more than the speed at which I could produce code. I began working with a repository, individual changes and their history. Instead of saying ‘fix this’, I needed to describe what should change and, just as importantly, what must remain untouched.

A proposal could now be reviewed, compared with the previous version and reversed if necessary. After manually moving files between my computer and a server, being able to trace an exact change felt almost luxurious.

When speed overtakes judgement

Codex can prepare a remarkably large change in a short time. But the number of lines produced says little about correctness. I discovered that defining requirements, boundaries and meaningful tests mattered more than admiring the amount of code generated.

I now separate a proposed change, a passing test and a successful deployment. That is not a pedantic distinction: each can fail for a different reason. Whenever a tool announces that it is finished, I want to know precisely what evidence supports that statement.

The developer becomes a coordinator

Working with AI means spending more of my time making decisions, framing tasks and reviewing outcomes rather than typing syntax. That is convenient, but also a little unsettling. If I delegate both the implementation and the definition of success, I risk losing my understanding of the system.

Codex is not a magical employee who can assume responsibility for a project. It is a powerful tool whose output still needs to be understood, tested and sometimes rejected. That was the most important lesson of leaving manual chat-guided development behind.

Next, I will look at how Agent finds information in its own knowledge base.