The model is no longer the most interesting part of a coding agent setup.
That sounds wrong if you only watch demos. The demo is always about the model. It reads the issue, writes the code, explains the diff, maybe even runs the tests. Clean screen recording. Nice ending. Everyone claps.
Real projects are messier. The hard part is not getting an agent to produce code once. The hard part is making that work repeatable, inspectable, recoverable, and boring enough that a team can trust it on a Tuesday when nobody has patience for another magical workflow.
That is why the wrapper around the agent is starting to matter more than the agent itself.
Chat was the first interface. It is not the final one.
