Chapter 12. What agentic AI is

Key points

• Until now you handed files over; from here, the AI operates files itself
• An agent runs plan → execute → verify on its own
• Precisely because it is powerful, caution and human checking are preconditions

From handing over to entrusting

This is the destination of the guide —

the boundary between "using AI" and "designing the work together with AI"

Until now, a human handed files to the AI. From here we enter a world where

the AI, inside your computer, reads files, moves across them and carries work forward by itself

This is called agentic AI.

Foundations talked about treating AI as an agent with a role. Agentic AI takes that further: something that actually moves its hands and operates files.

What makes an agent different

The difference from conventional AI is easiest to see as stages.

• Conventional chat — the human asks, the AI answers
• AI-assisted editors — the human works, the AI assists
• Agentic — the human states the goal, and the AI plans, executes and verifies by itself

Ask "read all the documents in this folder and produce a one-page summary of what they share", and an agentic AI will open each file in turn, read it and write the summary, as one continuous piece of work.

Powerful, so proceed carefully — the security crux

This matters particularly for therapists, so let me be careful about it.

Agentic AI holds permission to read, rewrite and operate files on its own. Convenient, and by the same token,

precisely because it is powerful, it needs caution

Agentic tools have in fact been found vulnerable to attacks that smuggle in instructions from outside (prompt injection, among others).

Which is exactly why the Chapter 2 principles matter again here.

• always back up important files before using it (be ready for an unintended rewrite)
• never let an agent touch folders containing patient information
• watch what the AI is about to do as you go; do not leave it running unattended
• practise first with harmless files and invented data

The right order is to take on Parts 1 to 4 properly, and touch agents once you can judge what to entrust and what to check yourself. That is why this comes last.

Chapter 12 summary

• Agentic AI runs plan, execute and verify on its own
• It holds permission to operate files, so it is powerful but needs caution
• Back up, check, practise — and touch it once the ground is solid