Updates from last week, I have iterated a little bit more on my “agentic development rule” for building a new project bit by bit by asking new features and improvements and reviewing them as I go and I am quite happy with the results! I built a couple of mini projects, that are still growing and the agent is able to keep the pace of development, write the tests and even write the design decisions in a log for documentation.

In my day to day job I am trying to introduce also more rules and workflows for a bigger project and I am starting to think that the model are quite capable of working with any code as long as the task is clear and the documentation or the context is enough to complete the task. In fact I am starting to say “If the model can’t do it, it is a code smell” Maybe I need to document the code better, or have a better test suite or specify the problem in a more detailed way…

I don’t like long rules that are difficult to review and I like bullet points because when I find something that the model does that I don’t like or that is going in the wrong direction, I add a bullet point for it and usually that is enough to correct the behavior.

See the comments for a link to my rule for building new features and a tetris clone project with several commits built only with agents (each commit is a task request for the agent that it did successfully)