The agent is instructed to use Laws of UX[0] when reviewing application flows and cite specific laws with changes to the builder agents. Look-and-feel in general still requires some tuning, but I have found that: 1) using a discrete agent in this role that only looks at the flow and is given the end-user context, 2) pointed to a set of guidance has really improved autonomy for UX.
It certainly is less important as the models get better at meta planning, but this approach is still very useful.
Often all you need to do is define a set of beliefs and values for the role and it will operate much better than rules or guidance.
Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Stream (https://arxiv.org/pdf/2606.01770) talks about this and the efficacy of role- and objective-based agents on unbounded streams of heterogeneous tasks.
I recently wrote down some of my own experience with this: LARPing with Agents (and Why It Works) (https://chrlschn.dev/blog/2026/10/larping-with-agents-and-wh...) after someone called out my role-based agents as "larping". Short of it: role-based, fully "JIT" agents have increased autonomy in my coding workflows end-to-end. I get very good outputs with few skills required and less human instruction input.
What's the best approach to applying these in practice? Should I evaluate every screen with each of these laws in mind or should I evaluate the app entirely one law at a time?
I have a dedicated agent profile that specifically owns this task and I tell the other agents to only send screenshots to it and ask for feedback.
Hoping Jev on open router can save me.