How to Create Powerful AI Agents that Work Together (Live Demo)

Key Concepts

• Agentic AI
• LLMs in a loop
• Orchestrator agent
• Writing agent
• Critiquing agent
• "Is it done" agent (Decision-making agent)
• Marketing plan generation
• Prompt engineering
• Iterative refinement

Agentic AI and LLMs in a Loop: Writing a Marketing Plan

Introduction

The video explores the concept of agentic AI, focusing on how Large Language Models (LLMs) can be used not just for writing, but also for making decisions and taking actions. The core idea is to create a loop of LLMs, where different agents interact to achieve a more complex goal than a single LLM call could accomplish. The example used is writing a marketing plan for a new product.

The LLMs in a Loop Model

The model consists of three main agents orchestrated to iteratively refine a marketing plan:

1. Orchestrator: This agent manages the entire process, calling on other agents and directing the flow of information.
2. Writing Agent: This agent is responsible for generating and revising the marketing plan based on the product description and feedback.
3. Critiquing Agent: This agent reviews the marketing plan and provides feedback on its strengths and areas for improvement.
4. "Is it done" Agent: This agent acts as a decision-maker, assessing whether the marketing plan is good enough to be finalized or if it needs further revision.

Step-by-Step Process

The process unfolds as follows:

1. The Orchestrator initiates the process by calling the Writing Agent to create an initial marketing plan based on the product description.
2. The Orchestrator then sends the plan to the "Is it done" Agent to determine if it's final.
3. If the "Is it done" Agent responds with "no," the Orchestrator sends the plan to the Critiquing Agent for feedback.
4. The Critiquing Agent provides a critique of the plan, highlighting areas for improvement.
5. The Orchestrator sends the original plan and the critique to the Writing Agent to create a revised version of the plan.
6. Steps 2-5 are repeated in a loop until the "Is it done" Agent determines that the plan is good enough to be finalized, responding with "yes."

Code Implementation

The video demonstrates a Python implementation of this model using Jupyter Notebook. The code utilizes two LLMs: OpenAI and X AI. The code defines three agents, each with a specific prompt:

• Writing Plan Agent: The prompt instructs the agent to write a marketing plan based on the product description. For revisions, the prompt includes the product description, the previous version of the plan, and the critique.
• Critiquing Agent: The prompt instructs the agent to provide a critique of the marketing plan, highlighting strengths and areas for improvement.
• Determine if Final Agent: The prompt asks the agent to assess whether the marketing plan is good enough to be final and to respond with either "yes" or "no" enclosed in <final> tags, along with a rationale.

The code then implements the loop, iteratively writing, critiquing, and revising the plan until the "Is it done" agent determines that it's final. The code saves each version of the plan for later analysis.

Results and Analysis

The video shows that the loop iterated three times before the "Is it done" agent determined that the plan was final. The speaker notes that each revision of the plan became more fleshed out, with richer competitive analysis in the final version. The formatting also improved over the iterations. The speaker concludes that the final version of the plan is a useful tool that can be drawn from.

Key Arguments and Perspectives

The video argues that agentic AI, where LLMs are used not just for writing but also for decision-making and action-taking, can enable more sophisticated workflows and allow software to do more for users. The LLMs in a loop model is presented as a simple example of how this can be achieved.

Notable Quotes

• (Referring to the "Is it done" agent) "...this was a very simple example of an agent that's a decision-making agent but in future videos I'll show you uh more elaborate examples..."
• "...agents and thinking of using Gen AI in an agent-based way this agentic AI style can allow you to do more sophisticated workflows more things on your own really enable uh and depend on the software to do more for you than just write you a document."

Technical Terms and Concepts

• Agentic AI: Using AI agents to make decisions and take actions, not just generate text.
• LLMs in a loop: A system where multiple LLMs interact iteratively to achieve a complex goal.
• Prompt engineering: Designing effective prompts to guide the behavior of LLMs.

Logical Connections

The video logically connects the concept of agentic AI to the specific example of writing a marketing plan. It demonstrates how different agents can be orchestrated to iteratively refine the plan, leading to a better final product than a single LLM call could produce.

Synthesis/Conclusion

The video effectively demonstrates the potential of agentic AI and LLMs in a loop. By using different agents for writing, critiquing, and decision-making, the system can iteratively refine a marketing plan, leading to a more comprehensive and useful final product. The "Is it done" agent plays a crucial role in determining when the plan is good enough, allowing the loop to terminate and preventing endless iterations. The speaker suggests that this is just a simple example and that more elaborate agent-based systems can enable even more sophisticated workflows.