By Vishal Deshpande, Chief Data Analytics Officer
Throughout our “AI in Plain English” series, we’ve explored examples and shared analogies to help non-experts better understand key AI concepts. For those of us supporting federal missions, understanding AI fundamentals is crucial for effective implementation and strategic planning. AI continues to progress at a rapid pace, bringing even more shifts. One powerful concept gaining traction: Agentic AI.
In this two-part blog, I’ll explain Agentic AI in easy-to-understand terms. In this first installment, I’ll describe the difference between a model and an agent. I’ll also explore how a work crew of agents can help address complex mission problems.
What’s a Model vs. What’s an Agent?
In the world of Gen AI, models and agents are very different. Let’s break down each.
Model (e.g., Large Language Model - LLM)
- A model is essentially a trained algorithm. Think of it as a highly sophisticated pattern recognition engine.
- It excels at tasks like generating text, translating languages, and answering questions based on the data upon which it was already trained.
- However, a model is passive. It responds to prompts but doesn't independently pursue goals or take actions in the real world.
- When a model hallucinates, it's because the model is generating an output based on patterns it has seen, even if those patterns don't reflect reality. (See our recent blog on hallucinations entitled “IDK is A-OK: Taming LLM Hallucinations.”)
Agent (Agentic AI)
- An agent builds upon a model by adding layers of autonomy and decision-making capabilities.
- It can:
- Plan: Break down complex tasks into smaller steps.
- Act: Use tools and APIs to interact with the real world.
- Reason: Evaluate information and make decisions.
- Learn: Remember past experiences and adapt its behavior.
- An agent is active. It can pursue goals and take actions to achieve them.
- By incorporating tool use, and verification steps, an agent is far less likely to hallucinate than a model used on its own.
Assembling Your Agentic AI Work Crew: Specialized Agents & Manager Agents
Now, let's take Agentic AI a step further. Imagine a complex government effort like disaster relief coordination. By nature, you’ll need a variety of talents—including some very specialized skills—to address this challenge. You’ll also need a level of management to orchestrate the activities of all those workers.
How would Agentic AI approach disaster relief coordination? Agentic AI would leverage multiple specialized agents, managed by a central manager agent, essentially forming a very targeted and specialized work crew. Let’s explore this further.
We’ll need a variety of Specialized Agents, specifically designed to execute very targeted tasks. For instance:
- A "data analysis agent" can process sensor data and social media feeds.
- A "communication agent" can draft and send alerts to citizens.
- A "resource allocation agent" can determine where supplies are needed most.
The activities of these specialized agents are orchestrated by a Manager Agent. The Manager Agent:
- Breaks down the overall task into smaller, manageable sub-tasks.
- Assigns sub-tasks to the appropriate specialized agents.
- Monitors the progress of each agent and provides feedback to refine tasks.
- Integrates the outputs of the specialized agents into a coherent whole.
- Handles conflict resolution between agents.
The benefits of this approach? Engaging multiple agents in this way reduces complexity. It allows for more robust solutions than a single agent could provide.
Agentic AI, though, is not without its risks. In part two of this blog, I’ll delve further into the advantages and disadvantages of Agentic AI. I’ll also share my thoughts on what Agentic AI adoption may look like in the federal landscape.
