How does the agent choose which tool to use?

The language model reads the tool names and descriptions along with your instructions and decides. Clear descriptions and instructions are the biggest lever you have over tool selection.

Can agents be unpredictable?

Yes. They make their own decisions, so test with a range of questions, limit what tools can do, and start with the simplest approach that solves your problem.

Langflow Module 4: Building an Agent with Tools

Introduction

A flow with a fixed path always does the same steps. An agent decides for itself which tools to use. Langflow's Agent component lets a language model call other components as tools. This module builds an agent from the Simple Agent template and shows how to give it new tools.

Prerequisites

Modules 2 and 3 completed
Familiarity with the idea of AI agents and tools

Step-by-Step Guide

Step 1: Start from the Simple Agent template

Create a new flow and choose the Simple Agent template. It contains an Agent component, Chat Input and Chat Output, and two tools: Calculator and URL. Check the Agent's Language Model field and choose a provider and model you enabled in Settings.

Step 2: Understand the Agent component

The Agent has a Language Model field, Agent Instructions for the system prompt, an Input for the user message, and a Tools port. It also has options such as Current Date, Handle Parse Errors, Verbose, and the number of chat history messages to include. Its outputs are Response and Structured Response.

Step 3: Write clear agent instructions

Tell the agent its role, which tools it has, and when to use them. Specific guidance beats generic text, because the agent picks tools based on the instructions and the tool descriptions.

You are a research assistant.
Use the Calculator for any arithmetic instead of computing in your head.
Use the URL tool to read a web page when the user gives you a link.
If a tool fails, say so and explain what you tried.

Step 4: Connect a component as a tool

To make a component available to the agent, enable Tool Mode on it, then connect its Toolset output to the Agent's Tools port. A component's description and field names tell the agent what the tool does, so keep them clear. You can attach several tools to one agent.

Step 5: Test and watch the tool calls

Open the Playground and ask something that needs a tool, such as a calculation or a request to summarize a URL. The Playground shows when the agent calls a tool and what it returned. Try questions that need no tool, too, to confirm the agent does not call tools unnecessarily.

FAQs

How does the agent choose which tool to use?: The language model reads the tool names and descriptions along with your instructions and decides. Clear descriptions and instructions are the biggest lever you have over tool selection.
Can agents be unpredictable?: Yes. They make their own decisions, so test with a range of questions, limit what tools can do, and start with the simplest approach that solves your problem.

Summary

The Agent component lets a model pick and call tools. Give it clear instructions, enable Tool Mode and connect Toolset to Tools to add capabilities, and test with and without tool-worthy questions.