Plans, and approving a plan

In chapter 2 the agent fixed the cracked printer one turn at a time, and nobody could see where it was heading until it got there. The support lead wants to know what the agent means to do before it does it, and a way to say no.

This chapter: How does an agent plan a case, change the plan when a result surprises it, and who gets to approve the plan?

You will learn

What a plan adds: the steps written down before any tool runs
Why a plan has to change when a result surprises the agent, and what that looks like
How approving a plan once replaces approving every call, and why a change goes back to the person
Why approval stops a planted instruction only if someone reads the plan

Words you’ll meet (4)

Plan: The steps the model writes down before it acts, one item per tool call. The host shows it on the clipboard.
Re-plan: The agent changing its plan part way, because a result told it something the plan did not expect.
Plan approval: A person approves the plan once before anything runs. A call outside the approved plan comes back to them.
Prompt injection: Instructions hidden in data the model reads, such as a customer email, written to make it do something else.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 10

The cracked printer again, and the printer is out of stock. In chapter 2 the agent picked each step as it went, so nobody could see where it was heading until it got there. Plan mode asks the model to write its steps down first, before it calls anything. The host puts them on the clipboard by the model. Turn plan mode on.