At level 4 the model stops being just a chat and starts using tools. They're the "hands" of the set-up around the model that chapter 4 talks about: a model on its own can only write; connected to tools it can search, read, calculate and act. This fact sheet lines them up, with what you risk using each one and the stage where you try it.
| Tool | What it does | When it's worth it | What to watch out for | Stage |
|---|---|---|---|---|
| Web search | Searches pages and cites them in the answer | Recent facts, news, prices | Sources that don't say what the model reports | 15 |
| Deep research | Works for minutes across many sources and returns a report | Broad questions, market or sector overviews | Long source lists, to be spot-checked | 15 |
| Connectors (MCP servers) | Connect the model to email, calendar, file stores, data services | Work that starts from your own data | Permissions that are too broad; every connected tool takes up space in the context window | 16 |
| Code execution | Writes and runs programs on files and data | Calculations, tables, charts | A sum done in its head instead of actually run | 4, 18 |
| Folder on your computer | Reads your files and writes new ones | Work on many documents | Actions that overwrite or delete | 18 |
| Browser driven by the model | Opens sites, reads, clicks, fills in forms, downloads | Data a site shows but won't let you export; sites with no direct connection | It acts with your logins; pages that contain hidden instructions for the model | 18 |
| Reusable instructions | A brief written once and called up when needed | Work that comes back the same each time | Old instructions nobody rereads | 14, 20 |
| Scheduled tasks | Start a task at a set time or when something happens | Periodic reports, recurring checks | Results nobody looks at anymore | 20 |
A word on connectors, because the technical name comes up often. Almost all of them go through an open standard, the Model Context Protocol: for each service there's a small program, the MCP server, that describes to the model the things it can do with that service, and the same server works with any tool that speaks the standard. A data provider can offer one too, so the model can query it directly. The book explains the mechanism in appendix Q, "How the Model Context Protocol works" (page 262), following the made-up but typical case of a market research provider.
The browser is the newest tool and the one to use with the most care. It means this: the model takes control of a browser, the one open on your computer or a separate one inside the app, and browses as you would: it opens the pages, reads them, clicks on menus, fills in the boxes, downloads the files. You watch it do so on screen, step by step, and can stop it whenever you want; stage 18 has the screenshot of one of these sessions. It's for when the data sits on a site that shows it on screen but offers no export or direct connection: an interactive table, a chart, a hundred pages of results. The model works inside your session, with your logins and your subscriptions, and should be treated like a colleague you've lent your computer to. The sites' terms of use apply to it too, and paywalls or anti-bot checks are not to be got around: in the instruction always write that when it hits a login request, a payment request or an anti-bot check it must stop and ask you.
In the book. Chapter 4, section 4.4 "The harness: the set-up around the model", page 89; chapter 6, section 6.2 "The four pillars of Agentic AI", page 105, for tools and the Model Context Protocol; appendix Q, page 262. The book on Amazon.