Tools and actions

From words to actions

On its own, a model can only produce text. A tool is a function the agent is allowed to call, such as "fetch this web page", "search the mailbox" or "create a calendar event". The model does not run tools itself. It writes a structured request ("call fetch with this address"), the platform checks the request, runs the tool, and hands the result back as text.

That split matters for safety. The model proposes, the platform disposes. Every limit that is enforced by the platform (which tools exist, what they may touch, whether a person must approve) holds no matter what the model was persuaded to attempt.

Kinds of tools

• Read tools fetch or look something up and change nothing: reading a page, searching mail, running a read-only query.
• Write tools change something: drafting, sending, creating, installing, deleting.
• Helper tools support the agent itself, such as saving a note to memory.

Read tools are low-risk and can usually run freely. Write tools that are hard to undo, such as sending a message to a customer, deserve a human check.

Least privilege

The most important design rule for tools is least privilege: give an agent only the tools it needs for its job, and give each tool only the access it needs. An agent that summarises web pages needs a fetch tool, not a mail tool. An agent that reads mail does not need permission to send it.

How this looks in PGB Agents

In Tools, an agent can only use what you tick, and the choice is saved with the agent. Tools are grouped: Web (http.fetch), PGB (a read-only database query), Marketplace, Memory, Notify, PGB Business (mail, calendar and drive) and Connected servers. Tools marked Needs approval pause the run and wait for you every time. Sending email, creating calendar events, emailing you and installing packages always need approval.

Try it yourself

1. Open "Concept lab" and look at Tools. Read the description of each tool in the Web, Memory and Notify groups.
2. Tick http.fetch and notify.email. Notice that only notify.email carries the Needs approval badge.
3. In the instructions add: "If the run input is a web address, fetch the page and explain its main topic in two sentences." Save.
4. Run it with the input https://example.com and open the run. In the trace find the step where the agent called http.fetch, and choose Show details to see the exact request.
5. Untick notify.email again. You do not need it for this course, and fewer tools means fewer things to go wrong.

Recap

• A tool is an action the model can request and the platform carries out.
• Read tools are low-risk, write tools need care, and the riskiest need a person's approval.
• Least privilege: the fewest tools, with the least access, that do the job.
• In PGB Agents, an agent can only use ticked tools, and some tools always Need approval.