Starting with a model key and a scaffold is how most agent projects begin, and it is also why so many stall. The quieter work in this post is the real work: naming the job, the data the agent may touch, which actions need a human before they run, and what "done" looks like when the demo is over.
Those four answers are the problem statement. Without them you get a chat wrapper that can read everything and change nothing safely, or the opposite: a script that can change everything with no review trail.
At Ascendra Ventures we treat that list as one atomic technology problem before any framework debate. If the team cannot write the job and the gates on one page, the build is not late. It has not started. Scaffolding after those answers land is usually a short project. Scaffolding before them is where the rewrite budget goes.