Why agents change the design problem
A chatbot answers. An agent acts. It sends the email, changes the record, books the slot. That moves the risk from "the answer was wrong" to "something happened that I did not want".
NN/g describes generative AI as a shift to intent-based outcome specification: the user states the result they want and the system decides how to get there. That reverses who is in control. Good agent design gives some of that control back, on purpose.
Show the plan before the work
Before an agent starts a multi-step task, show what it intends to do in plain steps. Let the user edit, remove, or approve them. This single pattern removes most surprises.
State the goal
Repeat the request in the agent's own words so misunderstandings surface early.
Propose the steps
List the actions in order, with the tools and data each one touches.
Ask where needed
Mark the steps that need approval, such as anything irreversible or external.
Run and report
Show progress as it happens and summarize what changed at the end.
Permissions that grow with trust
Do not ask users to grant broad access on day one. Start with read-only or draft-only behavior, and let people widen the scope as they see the agent behave.
| Level | What the agent may do | Good for |
|---|---|---|
| Suggest | Propose actions, never run them | First week of use |
| Draft | Prepare the work, wait for approval | Customer-facing changes |
| Act with review | Run, then show a summary to undo | Routine, reversible tasks |
| Act | Run without asking | Low-risk, well-understood tasks |
Design the handoff in both directions
Agents get stuck. When they do, they should say so, say why, and pass the work to the person with the context needed to continue. Do not make the user start over.
The reverse matters as much. Users should be able to step in mid-task, edit what the agent has done, and hand control back. Treat the work in progress as a shared document, not a black box.
Keep a log users can read
A readable activity log is the agent's receipt. It lists each action, when it ran, what it touched, and a way to undo it. It also helps your team debug and builds the trust that lets users delegate more.



