A good Agent doesn't call a tool right away. It writes a short plan first, and only then acts one step at a time.
Let's think about it this way for a moment
The plan is the route. Acting is the step. Skip the plan, and an Agent tends to read the wrong file, call things twice, and lose its way. Make the plan too long, though, and it burns time before any work even starts. For your first Agent, ask it to write only 3–5 steps. Show the plan to the user, and trust goes up. See "find the file, read it, pull out 3 points, stop" laid out plainly, and a person can catch a mistake and stop it before it happens.
Let's connect this to everyday life
A user asks, "what's important about this note?" A bad Agent immediately searches the web. A good Agent shows a plan first. 1) confirm the file name 2) read it 3) pull out 3 points 4) cite the source line 5) stop. Only once the plan is accepted does it start calling tools. This is a lightweight form of human-in-the-loop.
Let's try it out together
User goal: important points from notes.txt
Plan:
1. Confirm the file exists
2. Read notes.txt
3. Pick 3 points that are in the file
4. Show each point with a short quote
5. Stop
Wait for approval, then act.You'll be able to have an Agent write a plan that can be approved before it acts.5-minute try-it
Write a 5-step plan for "summarize today's dated note in this folder." Note at which step it needs human approval.
A quick word of caution
Write "search the internet if needed" into a plan, and the Agent will tend to slip past its permissions. If it's not allowed, leave it out of the plan entirely.