Give an agent a boundary people can see

The pitch for agents is that you hand a model some tools and let it find the route. In my experience the route is the product. If I can't say which tools it may call and what a person has to confirm, what I have is a demo that sometimes works.

I keep the boundary small on purpose. Read these documents, draft a reply against them, and stop before anything is sent, changed, or deleted. How many tools are attached matters much less than where that stopping point sits.

The person using it should be able to see the boundary too. What the agent read, what it proposed, and what someone approved belongs in the interface, not only in a log for engineers. People who can't see the line either trust the agent too much or stop using it.

MCP and tool calling are most useful to me when they reach into systems a team already has. If the work already happens somewhere, I'd rather the agent meet it there than invent a new place for it to live.

Before I ship one, I ask whether someone who has never seen the prompt could tell, from the screen alone, what the agent is allowed to do. If they couldn't, the boundary isn't finished.