Skip to main content

Dry run first

Most people start an agentic UI with an approval button. That comes too early. A user who has not yet seen anything useful has no reason to trust the agent, and a bare "Do you approve?" is a question they cannot answer.

Start with a dry run: the agent does the work as a draft, shows the result, and changes nothing yet. Only when something concrete is on the table, and the user can see its value, does the user decide.

The order of trust​

  1. Show the plan. The agent says what it understood and which steps it would take.
  2. Show the result as a draft. The user sees what would happen, with nothing executed.
  3. Let the user adjust. Change a step, remove an item, ask for another option.
  4. Ask for the decision. Now "yes, I want this" means something.
Agent: I found 12 invoices older than two years.
       Dry run: nothing has moved yet. Here is what I would archive.
       [ Show the 12 files ]   [ Remove some ]   [ Archive these ]

Why a dry run builds trust​

  • The user sees the process, step by step, before anything is at stake.
  • A wrong guess costs a glance, not a repair.
  • The decision at the end is informed, so approvals are given with care and not by habit.

While the agent then carries out the work, it keeps showing the steps; see streaming and progress.

Learning goals​

After this lesson you can:

  1. explain why an approval button at the start is too early;
  2. describe the four steps from plan to decision;
  3. design a dry run for an action that is hard to undo.