Let the models improve underneath the program

I wanted to write agent systems more declaratively, so I would not have to rebuild their control flow every time the models improved.

The familiar engineering ideas were still useful: composition over inheritance, inversion of control, and components whose dependencies could be supplied by the surrounding system.

The tension was between constraint and complexity. A prescriptive pipeline gives you control over the steps, but it can also keep a better model inside decisions made for an older one. Taking advantage of the improvement then means revisiting the pipeline.

My aim was to state the requirements in reusable components and let an intelligent system organize their execution. The components would still constrain the work; the system would have room to choose how to do it.

That also offered a way to treat an organization’s process knowledge as code: something it could version and improve over time.