Configure the Tool Calling Agent#

Configure the NVIDIA NeMo Agent toolkit tool calling agent as a workflow or a function.

Requirements#

The tool calling agent requires the nvidia-nat[langchain] plugin, which can be installed with one of the following commands.

  • If you have performed a source code checkout:

uv pip install -e '.[langchain]'

  • If you have installed the NeMo Agent toolkit from a package:

uv pip install "nvidia-nat[langchain]"

Configuration#

The tool calling agent may be utilized as a workflow or a function.

Example 1: Tool Calling Agent as a Workflow to Configure config.yml#

To use the tool calling agent as a workflow, configure the YAML file as follows:

workflow:
  _type: tool_calling_agent
  tool_names: [wikipedia_search, current_datetime, code_generation]
  llm_name: nim_llm
  verbose: true
  handle_tool_errors: true

Example 2: Tool Calling Agent as a Function to Configure config.yml#

In your YAML file, to use the tool calling agent as a function:

function_groups:
  calculator:
    _type: calculator
functions:
  math_agent:
    _type: tool_calling_agent
    tool_names: [calculator]
    llm_name: agent_llm
    verbose: true
    handle_tool_errors: true
    description: 'Useful for performing simple mathematical calculations.'

Configurable Options#

  • workflow_alias: Defaults to None. The alias of the workflow. Useful when the Tool Calling agent is configured as a workflow and need to expose a customized name as a tool.

  • tool_names: A list of tools that the agent can call. The tools must be functions or function groups configured in the YAML file

  • llm_name: The LLM the agent should use. The LLM must be configured in the YAML file

  • verbose: Defaults to False (useful to prevent logging of sensitive data). If set to True, the agent will log input, output, and intermediate steps.

  • handle_tool_errors: Defaults to True. All tool errors will be caught and a ToolMessage with an error message will be returned, allowing the agent to retry.

  • max_iterations: Defaults to 15. The maximum number of tool calls the agent may perform.

  • return_direct: Optional list of tool names that should return their output directly without additional agent processing. When a tool in this list is called, its response is returned immediately to the user, bypassing the agent’s reasoning step.

  • description: Defaults to “Tool Calling Agent Workflow”. When the agent is configured as a function, this config option allows us to control the tool description (for example, when used as a tool within another agent).

Step-by-Step Breakdown of a Tool-Calling Agent#

  1. User Query – The agent receives an input or problem to solve.

  2. Function Matching – The agent determines the best tool to call based on the input.

  3. Tool Execution – The agent calls the tool with the necessary parameters.

  4. Response Handling – The tool returns a structured response, which the agent passes to the user.

Example Walkthrough#

Imagine a tool-calling agent needs to answer:

“What’s the current weather in New York?”

Single Step Execution#

  1. User Query: “What’s the current weather in New York?”

  2. Function Matching: The agent identifies the get_weather(location) tool.

  3. Tool Execution: Calls get_weather("New York").

  4. Response Handling: The tool returns 72°F, clear skies, and the agent directly provides the answer.

Since tool calling agents execute function calls directly, they are more efficient for structured tasks that don’t require intermediate reasoning.

Limitations#

The following are the limitations of tool calling agents:

  • Requires an LLM that supports tool calling or function calling.

  • Does not perform complex reasoning and decision-making between tool calls.

  • Since it uses the tool name, description, and input parameters, it requires well-named input parameters for each tool.