The real problem
An agent demo can impress in five minutes. But when you put it to work with real repositories, tools and permissions, problems appear that no prompt can solve.
The approach
I prefer to separate the execution runtime from the interface. The runtime owns sessions, tools, limits, and state; the UI displays intent, progress, evidence, and decisions awaiting approval.
Decisions that matter
The system must distinguish observation from modification. Reading a file is not the same risk as running a command or changing production. Authorization and limits must live outside model-generated text.