What Are Tools?

Tools are functions your agent can call to interact with the world. They’re the agent’s hands and eyes.

Without tools:

User: "What files are in ./notes?"
Agent: "I can't access your filesystem, but you could run 'ls notes' to check."

With tools:

User: "What files are in ./notes?"
Agent: [calls list_files("./notes")]
Agent: "You have 3 files: 2026-01-20.md, 2026-01-21.md, 2026-01-22.md"

Tools turn agents from advisors into actors.

Our Four Tools

We’ll implement four tools. Two are safe (read-only), two need approval (side effects):

Safe Tools (No Approval Needed)

1. list_files(directory)

• Lists files in a directory
• Read-only, no side effects
• Safe to run automatically

2. read_file(path)

• Reads content from a file
• Read-only, no side effects
• Safe to run automatically

Risky Tools (Approval Required)

3. write_file(path, content)

• Writes content to a file
• Side effect: Changes disk state
• Risk: Could overwrite important data
• Needs approval

4. send_message(channel, text)

• Sends a message (mocked in this tutorial)
• Side effect: External communication
• Risk: Could spam or leak information
• Needs approval

Tool Schemas: The Contract

Tools need schemas so the model knows:

• What the tool does
• What parameters it takes
• What types those parameters are
• Which parameters are required

We use JSON Schema format (OpenAI’s function calling standard):

Implementing Tool Schemas

Let’s define all four tools:

Interactive: Match Tools to Categories

Drag each tool to the correct category:

Implementing Tool Functions

Now let’s implement the actual Python functions that execute these tools:

Tool Dispatch: Routing Calls to Functions

We need a dispatcher that maps tool names to Python functions:

Approval Gate Logic

Finally, let’s implement the approval gate:

Visualizing Tool Execution Flow

Interactive Diagram

This interactive diagram requires JavaScript to be enabled.

Steps:

1. Agent proposes a tool call
2. Check if tool needs approval
3. If risky, ask user for approval
4. Execute tool function
5. Return result to agent

Try It: Add a New Tool

Exercise: Add an extract_todos(text) tool that finds TODO items in text.

Requirements:

• Tool name: extract_todos
• Parameter: text (string)
• Returns: List of TODO items found
• Safety: Read-only (no approval needed)

Hint: Look for lines containing “TODO”, ”- [ ]”, or ”## Tomorrow”

Quick Knowledge Check

Key Takeaways

You now know how to:

1. ✅ Define tool schemas using JSON Schema format
2. ✅ Implement tool functions that return structured results
3. ✅ Distinguish safe vs risky tools based on side effects
4. ✅ Build a tool dispatcher to route calls to functions
5. ✅ Add approval gates for operations with side effects

What’s Next?

In the next page, we’ll build the agent loop—the core logic that ties everything together. You’ll see how the model proposes tools, your code executes them, and results flow back to the model.