Inside the Mind: Tool Selection

In Module 9, we defined our tools. In this lesson, we look at how the Agent actually picks them. This process is not a "Lookup table"—it is a Semantic Choice made by the model.

1. The Prompt Injection

When you initialize an agent, LangChain creates a giant "System Prompt" that includes:

1. Your instructions.
2. A list of all available tools.
3. Their names and descriptions.
4. Instructions on how to output the JSON for the tool call.

---

2. Dealing with Hallucinated Tools

Sometimes an agent gets confused and tries to call a tool that doesn't exist (e.g., call_mom()).

• LangChain's fix: The system catches the error and sends a message back to the LLM: "Error: Tool 'call_mom' does not exist. Available tools are: [multiply, search]."
• The LLM then "Realizes" its mistake and picks a valid tool.

---

3. Visualizing the Selection Flow

graph TD
    M[Model] -- "Thinking" --> Choice{Which tool?}
    Choice -- "Match Description" --> T1[Tool 1: Statistics]
    Choice -- "Match Description" --> T2[Tool 2: Search]
    T2 -- "Tool Call" --> Exec[Python Execution]
    Exec -- "Data" --> M

---

4. Forced Tool Use

Sometimes you want the agent to Always use a specific tool (e.g., if you are building a specific DB-interface bot).

• You can set the tool_choice parameter to "Force" the agent to call your tool on every request.

---

5. Engineering Tip: Argument Quality

LLMs are prone to sending the wrong "Data Types" to tools (sending "Five" instead of "5").

• As we learned in Module 9, use Pydantic in your tool definitions to validate these arguments before the code runs. This prevents your server from crashing and gives the agent a clean "Validation Error" to fix.

---

Key Takeaways

• Prompt descriptions are the only way the agent knows which tool to use.
• The agent is re-invoked after every tool call to process the result.
• Tool hallucination is handled by automated error-correction messages.
• Validation is critical for maintaining the agent loop's stability.