Agentic AI in Action: How LLMs Handle Multi-Step Tool Requests

Key Concepts

• Agentic AI: Using GenAI agents to make decisions and call tools (APIs, proprietary data, Python code) to affect the outside world.
• Tool Calling: The ability of LLMs to interact with external tools to perform specific actions.
• Model Context Protocol (MCP): A framework for managing the interaction between the LLM and the tools.
• Task Planning and Execution: The ability of GenAI systems to determine the necessary steps and execute them to complete a task.

Agentic AI and Tool Calling: A Practical Example

The video discusses Agentic AI and its ability to use GenAI systems to make decisions and call tools to interact with the outside world. The core idea is to move beyond simple text generation and enable AI to perform actions that affect the real world.

HR System Example: Emergency Contact Information

The speaker presents a practical example of using Agentic AI in an HR system. The scenario involves a company that has merged multiple HR systems, resulting in inconsistent and incomplete employee data, specifically missing emergency contact information.

• Problem: Employee data is fragmented and incomplete, with missing emergency contact information.
• Traditional Solution: A junior HR person would manually review employee data, email employees, and update the database.
• Agentic AI Solution: An Agentic AI system can automate this process.

Step-by-Step Process

The speaker outlines the following steps for the Agentic AI system:

1. Prompt: A detailed instruction is given to the AI system, such as "Get emergency contact info for everyone." The prompt includes instructions to look up information in the database, email employees if the information is missing, update the database upon receiving a response, and follow up with non-responsive employees.
2. Tool List: A list of available tools is provided to the AI system, including:
   • List all employees (API call)
   • Look up information on one employee
   • Write and send an email to an employee
   • Read an email response from an employee
   • Update the HR database
   • Tell me when you're done (task completion signal)
3. LLM Interaction Loop:
   • The LLM receives the prompt and the list of tools.
   • The LLM determines the next action and requests a specific tool with parameters (e.g., "Call the list all employees tool").
   • The agentic AI system calls the requested tool and sends the response back to the LLM.
   • This loop continues until the LLM indicates that the task is complete.
4. Tool Execution Loop:
   • The agentic AI system parses the LLM's response to identify the requested tool and parameters.
   • The system calls the appropriate internal tool (API, database query, etc.).
   • The system packages the tool's response and sends it back to the LLM.

Key Arguments and Perspectives

• Feasibility: The speaker asserts that LLMs are capable of determining the next appropriate action in a task like this.
• Frameworks: Frameworks exist to facilitate this process, such as the Model Context Protocol (MCP).
• Potential: Agentic AI has the potential to automate and augment tasks, freeing up humans from grunt work.

Notable Quotes

• "Agentic AI is basically elevating Genai from just writing a response to a prompt... and actually getting things done like a human would get something done."
• "That's where the real potential is to really accelerate and augment us in doing some of the potentially grunt work of of um uh of work."

Technical Terms and Concepts

• Agentic AI: AI systems that can make decisions and take actions in the real world.
• LLM (Large Language Model): The AI model that processes the prompt and determines the next action.
• Tool: An API, function, or other resource that the LLM can call to perform a specific action.
• Prompt: The instruction given to the LLM, outlining the task and providing context.
• Parameters: The inputs required by a tool to perform its function.
• Model Context Protocol (MCP): A framework for managing the interaction between the LLM and the tools (mentioned for future discussion).

Logical Connections

The video logically connects the concept of Agentic AI to a practical HR system example. It breaks down the process into manageable steps, highlighting the interaction between the LLM and the tools. The speaker emphasizes the potential of Agentic AI to automate tasks and augment human capabilities.

Synthesis/Conclusion

The main takeaway is that Agentic AI, through tool calling, has the potential to significantly impact how tasks are performed. By enabling LLMs to interact with external tools, Agentic AI can automate complex processes, augment human capabilities, and ultimately increase efficiency. The HR system example demonstrates a real-world application of this technology, showcasing its potential to streamline operations and improve data management. The speaker plans to elaborate on frameworks like MCP in future videos.