Tool calling is a powerful feature that allows you to call external tools that you define and integrate them into your agent’s responses.
Instead of requiring you to build static, hardcoded workflows, tool calling enables you to build agents that dynamically choose and orchestrate the right combination of tools based on the specific context of each request.
This document describes why tool calling matters and how it works behind the scenes.
Jump to tool calling in Agent Builder to learn how to use tool calling in Agent Builder.
Traditional approach limitations: With pre-defined API workflows, you’d need to hardcode every possible scenario: “If customer mentions shipping, call shipping API. If they mention returns, call returns API.” This creates rigid, brittle systems that can’t handle the complexity of real customer inquiries.
Tool calling’s dynamic intelligence: The model acts as an intelligent orchestrator, analyzing the customer’s specific problem and dynamically selecting the most relevant tools. It handles parameter extraction, API sequencing, and data synthesis automatically.
Enables advanced reasoning patterns: Tool calling naturally supports ReAct (Reasoning and Acting) workflows, where the model can think through problems step-by-step, take actions by calling tools, observe the results, and then reason about what to do next.
A customer writes to your support team:
“I ordered two items last week but only received one and the item I got isn’t working.”
Your Agent Builder agent takes a customer’s request and uses tool calling to dynamically select the most relevant tools to use. Available tools might include:
get_order_details(order_id)
check_shipping_status(tracking_number)
initiate_return_request(order_id, reason)
check_inventory_status(product_id)
create_support_ticket(customer_id, issue_type)
Given the tools provided to the model, it can review the user’s request and automatically:
get_order_details()
to understand what was orderedcheck_shipping_status()
for the missing itemWhile the AI model intelligently decides which tools to use and what parameters to call them with, the model doesn’t actually execute the tools itself. Instead, there’s a two-step handoff process:
get_order_details
with order ID #12345”The AI then uses the results of the tool calls to continue reasoning and provide the final response.
Think of it like a project manager coordinating subcontractors: The AI project manager understands the entire project scope and knows exactly which specialists to engage and what work needs to be done. When a customer has an issue, the AI says “I need the shipping contractor to check tracking #12345, then the inventory contractor to verify stock levels.” But the AI doesn’t do the actual work; each subcontractor (your systems) has their own expertise and tools to execute their piece, then reports back to the project manager who coordinates everything into the final deliverable.
This architecture provides several benefits:
Now that you understand why tool calling matters and how it works behind the scenes, learn how to use tool calling in Agent Builder.