An agent needs an edge.

Define where judgment returns

When designing an agent workflow, I would start by drawing its boundary. What can it read? What can it change? Which decisions require a person? A broad instruction to help is a useful conversation starter, but an operating workflow needs more precise permissions.

I also want a clear stopping condition. The agent should know when the intended outcome is achieved, when it needs input, and when another attempt would repeat the same failure. Those edges make autonomy easier to trust because the people using the system can anticipate how it will behave.