Agenten-Kompass
Guide overview

ToolsPart 4 of 5 6 minutes

How does an agent carry out an action?

The model proposes a tool call. The application checks inputs and permissions, executes permitted calls and returns the result or an error.

One tool call, step by step

A calendar leads via an orange approval button to a robot arm placing an entry in a grid. A check mark at the end stands for the returned result.1Tool2Call3Check and execution4Result
The orange button stands for the check before execution. AI-generated illustration.

  1. Describe the interface

    Name, purpose and input schema explain which function can be called with which values.

    With LeaThe invitation tool expects a recipient and a start time.

  2. Propose the call

    The model selects a tool and generates the required arguments.

    With LeaThe model generates: invitation to Lea, Thursday, 9:30 am. Nothing has been sent yet.

  3. Check and execute

    Application and service check inputs and permissions and execute the permitted call.

    With LeaYou approve sending. The tool then creates the appointment.

  4. Return the result

    The result or error message returns to the run and can change the next step.

    With LeaThe returned appointment ID shows that the entry exists.

TRY IT YOURSELF

Read or change?

This distinction helps you decide where you would require approval. What does the tool do each time?

Search for free slots in an approved calendar

Reads The search changes nothing. Access to the data still has to be permitted.

Send the invitation to Lea

Changes A message leaves your system and can hardly be recalled. Approval makes sense here.

Search project notes for a decision

Reads The tool only returns a match. The notes stay unchanged.

Delete an existing appointment

Changes This changes the calendar and may affect others. Confirmation fits here too.

TRUE OR FALSE?

Three things people often say.

If a tool is reachable, every call is permitted.

False A tool being reachable and being permitted for a specific call are separate requirements.

MCP decides who may do what.

False The protocol creates a shared interface. Application and service enforce access rights and approval rules.

An error message can help the agent.

True It can help correct parameters or choose another permitted route.

Explore the mechanisms

A call is a handoff

A tool call passes a function name and inputs to executing software. A calendar search might include a time range and calendar identifier. Its result is then returned to the model. This keeps the model proposal, actual execution and observed result distinguishable.

OpenAI: Function calling
Authorize reading and changes deliberately

A search reads information; sending a message changes external state. Permissions should fit the task. Applications can require confirmation for sensitive actions. Tool availability and authorization for a specific call are separate prerequisites.

MCP specification: Tools
MCP standardizes connections

The Model Context Protocol describes exchanges between an application and connected services. Tools offer executable functions, resources supply context data and prompts provide reusable templates. The protocol supplies a common interface; the application and service continue to enforce permissions and approval rules.

MCP: Architecture overviewMCP specification: Tools
Errors are part of feedback

A tool may report invalid inputs, missing permissions or a service outage. This feedback helps correct parameters or choose another permitted route. The result must establish whether the requested change actually happened.

Anthropic: Handle tool calls

Fundamentals and primary sources reviewed on 5 November 2026. The examples explain the principle; specific products may differ.