Title: How to Create Powerful AI Agents That Work Together (Live Demo)
Resource URL: https://www.youtube.com/watch?v=mlL3ofEzBK8
Publication Date: 2025-04-22
Format Type: Video
Reading Time: 13 minutes
Contributors: Don Woodlock;
Source: Don Woodlock (YouTube)
Keywords: [Artificial Intelligence, Workflow Automation, Agentic AI, LLM Decision Agents, AI Marketing Plan Generator]
Job Profiles: Academic/Researcher;Machine Learning Engineer;Artificial Intelligence Engineer;Data Analyst;Chief Technology Officer (CTO);
Synopsis: In this video, Intersystems' Don Woodlock explains how using agents in a loop with large language models (LLMs) can create iterative improvements for complex tasks like writing a marketing plan.
Takeaways: [LLMs can be programmed not just to generate text but also to make decisions and take actions within workflows., Orchestrating multiple agents—such as a writing agent, a critiquing agent, and a decision agent—can automate iterative refinement., Incorporating a decision-making agent enables automated quality checks before finalizing outputs., Iterative loops involving critique and revision lead to more detailed, better-structured deliverables over time., Designing agentic workflows can move AI usage beyond single-call responses into complex, autonomous processes.]
Summary: The speaker introduces the concept of agentic AI, highlighting how agents can extend the capabilities of large language models (LLMs) beyond simple writing tasks. A structured workflow is presented involving three types of agents: a writing agent that drafts a marketing plan, a critiquing agent that provides feedback, and a "determine if final" agent that decides whether the draft meets quality standards. An orchestrator manages these agents, putting them into an iterative loop where the plan is continuously critiqued and revised until deemed final. 

Using a simple Python script with OpenAI and X.AI models, the video demonstrates how this setup improves output quality over multiple iterations, showing tangible enhancements in formatting, competitive analysis, and overall content richness. This approach underlines the potential of agentic AI to manage more sophisticated tasks autonomously, laying a foundation for future applications in more complex workflows.
Content: ## Introduction to Agentic AI

This overview examines the concept of agentic artificial intelligence (AI)—the practice of composing and orchestrating multiple large language model (LLM) agents to tackle complex, multi-step tasks. Rather than relying on a single prompt-and-response interaction, agentic AI leverages specialized agents that write, critique, decide, and iterate until a satisfactory outcome is achieved.

## Capabilities of LLM-Based Agents

LLM-powered agents can perform three primary functions:

1. **Content Generation**: Crafting text such as marketing plans, scripts, poems, code, and more.
2. **Decision Making**: Evaluating outputs against defined criteria to guide subsequent actions.
3. **Action Execution**: Triggering further processes or prompting other agents based on decisions.

While most users are familiar with the generative potential of chatbots for writing tasks, the decision-making and action-orchestration features unlock more sophisticated workflows.

## Designing a Multi-Agent Workflow

A practical way to illustrate agentic AI is through the process of creating a marketing plan for a new product. The workflow comprises four collaborating agents overseen by a central orchestrator:

### 1. The Orchestrator

This supervisory agent manages the entire sequence. It issues prompts to subordinate agents, receives their outputs, and determines the next steps in the loop.

### 2. Writing Agent

Responsible for drafting or revising the marketing plan. On the first iteration, it receives only the product description and generates an initial plan. On subsequent passes, it also ingests the previous plan and accompanying critique.

### 3. Critique Agent

Analyzes the latest marketing plan in conjunction with the product description and provides detailed feedback, highlighting strengths, gaps, and suggestions for improvement.

### 4. Finality Agent

Assesses whether the current draft meets the criteria for a final, company-ready version. It returns a simple “yes” or “no” (embedded within designated tags) and supplies a rationale to guide model performance, even though the rationale itself is not used downstream.

## Workflow Loop for Marketing Plan Creation

1. **Initial Draft**: The orchestrator invokes the writing agent with the product description to obtain the first draft.  
2. **Finality Check**: The finality agent evaluates the draft. If it returns “yes,” the process ends.  
3. **Critique**: If the draft is not final, the orchestrator solicits the critique agent’s analysis.  
4. **Revision**: The orchestrator instructs the writing agent to produce a revised draft, feeding it both the previous plan and the critique.  
5. **Repeat**: Steps 2–4 continue in a loop, with a predefined upper limit on iterations (e.g., ten), until the finality agent signals completion.

## Implementation in Python

A prototype was developed within a Jupyter notebook, leveraging two distinct LLM endpoints—one from OpenAI and another from an alternative provider. Key elements included:

- **Agent Definitions**: Prompt templates and system messages tailored for each agent’s role (writing, critique, finality).  
- **Loop Control**: A Python loop executed the orchestrator’s logic, storing each plan iteration, invoking the finality agent, and breaking if the plan was approved or the iteration limit was reached.

By modularizing the prompts and responses, the code maintained clarity and allowed independent tuning of each agent.

## Observations and Outcomes

After executing the loop, three distinct versions of the marketing plan emerged:

- **First Draft**: A basic outline with broad strategies.  
- **Second Draft**: Improved formatting and structure, with expanded sections on positioning and messaging.  
- **Third Draft**: Enhanced competitive analysis, richer detail on target segments, and refined go-to-market tactics.

The finality agent approved the third iteration, demonstrating the system’s capacity to iteratively refine content without direct human intervention at each step.

## Conclusion and Next Steps

This simple example highlights how agentic AI—combining generative, evaluative, and decision-making agents—can automate multi-stage workflows and yield higher-quality outputs than single-call LLM interactions. Future work may incorporate additional agent types (e.g., data retrieval, performance monitoring) and more elaborate decision logic to support a wider range of business processes.

## Feedback and Engagement

Your insights and suggestions are welcome. Please share feedback or topic ideas for future videos in the comments below. I review all submissions and look forward to continuing this exploration of agentic AI.