Agent Tool Calling: Getting Agents to Actually Do Things
This course walks through the full mechanism of agent tool calling (tool use / function calling): exactly which fields travel between a tool-call request and its response, the risk boundaries of the five common tool types (read, write, execute, search, call an external API), how to write a tool interface the model both picks correctly and fills correctly, and how to put permission tiers and prompt-injection defenses into practice. It's for developers who can already write basic prompts and have used a tool like Claude Code, but haven't yet studied the underlying tool-calling protocol systematically. It does not teach any specific agent framework (LangChain, AutoGPT) and does not cover model training or fine-tuning — it focuses on the chain itself: how the model requests an action and how the host safely executes it.
Course outline
- From "Just Talking" to "Taking Action": Why Agents Need Tools
- The Full Round-Trip of a Tool Call
- Five Common Tool Types: Read, Write, Execute, Search, Call
- Designing Tool Interfaces: Name, Description, Parameters, Return Value
- Permissions and Safety: The Boundaries of What an Agent Can Do
- Hands-On: Wiring Three Tools onto an Agent
Learning goals
By the end of this course you'll be able to:
- Articulate the "the model only proposes, the host executes" line in tool calling, and use it to judge whether a task actually needs tools at all
- Read and hand-write a full tool_use / tool_result round-trip, including the batched return of several parallel calls
- Classify the five tool types (read, write, execute, search, call an external API) by blast radius and spot the trap each one is most prone to
- Write the tool description, JSON Schema, and return value that let the model pick the right tool, fill the parameters right, and self-correct after a failure
- Grade tool operations with allow / ask / deny, and recognize the risks of over-authorization and the lethal-trifecta combination
- Build, from scratch, a tool-calling agent with an execution loop, a registry, and a safety valve
Prerequisites
- You can write basic prompts and understand the basic shape of a conversation with an LLM
- You've used Claude Code or a similar AI coding tool and know it can read/write files and run commands
- You can read basic JavaScript / Node.js code (lesson 6 has you follow along in code)
- No machine-learning or model-training background required
Estimated time
About 3-4 hours, including the hands-on exercise in each lesson.