Hallucinated Actions: The Ghost in the Machine

A "Normal" hallucination is when an LLM lies about a fact (e.g., "The capital of Mars is Elonville"). An Agentic Hallucination is when the LLM lies about its own capabilities. It tries to use a tool that does not exist.

1. What a Hallucinated Action looks like

You give the agent a Search tool and a Write_File tool. The Agent's output: "Thought: I have the data. Now I will send an email to the customer. Action: send_email Action Input: {'to': 'user@example.com', 'body': 'Hello!'}"

The Error: Your code crashes because you do not have a send_email function defined.

2. Why Agents "Make Up" Tools

Training Data Bias: Models like GPT-4 have seen millions of lines of code. They "know" that a send_email or get_database_row function usually exists in a programmer's world.
Over-Helpfulness: The model wants to reach the final goal so badly that it "Invents" a shortcut to get there.
Ambiguous Interface: If you don't use "Native Function Calling," the model is just writing text. It doesn't "know" it's restricted to a specific list unless the prompt is very strict.

3. Detecting Ghost Tools

Your "Control Loop" (Module 2) must be defensive.

def control_loop(agent_choice):
    available_tools = ["search", "calculator"]
    
    if agent_choice["action"] not in available_tools:
        # 1. Stop the execution
        # 2. Feed an error BACK to the agent
        return f"Error: The tool '{agent_choice['action']}' does not exist. Please only use: {available_tools}"

4. Preventing Hallucinated Actions

A. Use Smaller, Specialized Toolboxes

Don't give one agent 50 tools. Give it 5. The more choices you give an LLM, the higher the "Interference" between tool names, and the higher the chance of it hallucinating a hybrid tool.

B. The "Available Tools" Reminder

In every turn of the chat history, remind the model of its limits: "Current State: You are in a restricted environment. You have 3 tools: [A, B, C]. Any other tool call will result in a system failure."

5. Visualizing the Validation Gate

graph TD
    LLM[Agent Output] --> Gate{Is Tool in Registry?}
    Gate -- Yes --> Run[Execute Code]
    Gate -- No --> Feedback[Tell Agent: Tool Unknown]
    Feedback --> LLM

Key Takeaways

Hallucinated Actions are attempts to use non-existent tools.
They are driven by the model's prior training on generic codebases.
Always implement a Validation Gate in your control loop.
Narrow toolkits are significantly more reliable than broad ones.

Module 5 Lesson 3: Hallucinated Actions