The agentic approach in AI moves beyond a single, potentially complex API call by employing a team of specialized “agents.” Think of it as replacing a Swiss Army Knife with a dedicated set of tools. These agents aren’t monolithic; they can leverage different Large Language Models (LLMs), fine-tune parameters like temperature for varied creativity, and access specific tools and data sources relevant to their function.
However, simply having multiple agents isn’t enough to constitute a true agentic system. The key ingredient is cooperation.
Consider a familiar scenario: a contact center. Instead of a single system handling everything, an agentic approach might work like this: First, an “intake agent” categorizes each incoming email (commercial, technical, or general). Next, a “sentiment analysis agent” assesses the customer’s emotional tone. Emails flagged as highly negative are immediately escalated to a human agent for personalized attention. The remaining calls are then routed to specialized “commercial,” “technical,” or “general” agents based on the initial categorization.
Each of these agents could have access to specific tools, like a product catalog for the commercial agent or a troubleshooting database for the technical agent.
Using a popular agentic framework like LangGraph, this defined workflow can be visualized as a network of interconnected agents:
Notice how the relationships between agents are explicitly defined, specifying who can communicate with whom.
Furthermore, each agent can be tailored with unique characteristics. Let’s take a look at a few of them:
The beauty of this approach lies in the flexibility of organizing these agent relationships in numerous ways to suit different tasks:
What I’ve described so far are the fundamental building blocks. Now, imagine taking this a step further. What if you didn’t want to predefine the exact process every time? This is where the concept of a central “orchestrator agent” comes in. Its role is dynamic: to analyze the incoming task, devise a suitable strategy using the available agents, distribute specific jobs, and even determine if a task has been completed successfully.
This orchestrator-driven model is the foundation of more autonomous agentic systems.
Currently, there are over ten prominent agentic frameworks, each offering varying degrees of agent autonomy.
For example, while LangGraph emphasizes a structured, pre-defined process, a framework like AutoGen positions the orchestrator as the central figure, dynamically coordinating the work of other agents. This is particularly advantageous when you need your agents to adapt to diverse tasks and avoid rigid adherence to a single workflow.
On the other hand, if consistency and strict adherence to a process are paramount, LangGraph’s “follow-the-process” approach might be more suitable, ensuring agents remain focused and less prone to “creative” deviations.
Leave a Reply