Agentic AI is all the new craze in the tech world. I am intrigued with the notion that a AI agent will be able to access specific part of a system and learn from the tasks the system performs. Over the time it will learn and adapt based on the interactions.
What could possibly go wrong?
The answer is a lot can go wrong but we can't stop working in the fear that things might go wrong. Nothing would get accomplished.
The right question to ask is -
What are AI Agents and what are the kind of problems they will solve for us?
Agentic AI or AI Agents are systems that use LLMs to autonomously accomplish specific tasks in coordination with third party systems with a goal of having impact on a specific use case.
May be this IBM Definition might seem more formal.
The way these AI Agents are meant to work depends on the initial goal and eventual objective of the use case. It can be a repetitive automation that enables a team to automate the mundane task and gain valuable insights over a period of time or it can be a sophisticated system that learns over time from the tasks it is performing to improve on itself. The use cases can be wide ranging but the goal and the return needs to be agreed upon in advance.
As it is with every autonomous system, there are benefits and risks that accompany with it. For teams building such AI Agents, the challenge will be to balance the benefit across risk & costs.
Are there any AI Agents already in the Market?
There is a case that the LLMs in the market are agents in themselves. A general purpose agent that replaces search engine for lot of people. If we are to go with the definition as mentioned above then the Agentic AI should be use case specific and more purpose driven.
The good folks at MIT have an index to maintain a database with the list of AI Agents currently in the market -
Database of AI Agents
Their criteria to be classified as a Agent is mentioned in the diagram below -
I am sure the criteria for an AI Agent to be included in the database will be broadened or improved upon as new evidence is learned from the ever expanding list of AI Agents.
Meanwhile, there are larger questions on hand regarding the standardization of the process of building AI Agents and a way for them to communicate with each other.
Let's be honest, what is the point of building AI Agents if they all work in silos? Hold off on answering that for subsequent research and exploration of AI Agents in the field.
The question I am more interested in right now is -
How can we standardize the communication of AI Agents with the rest of the application ecosystem?
A open standard called Model Context Protocol (MCP) enables developers to build a secure, two-way connections between data sources or third party applications and the AI systems or LLMs.
At its core, MCP follows a client-server architecture where a host application (Data sources and internal applications) connect to a MCP server and output the result to a third party system or back to the host application. The entire communication is done using APIs.
It is oversimplified explanation of how MCP works but you can see at its core its is a simple software development architecture with enhanced capability due to LLMs.
More details can be found here.
As it is with Software Development Architecture, even MCP has different elements of its architecture modularized by service providers to make it easy to get started with building AI Agents.
I will like to work with different SDKs to understand the architecture in practice. May be a good topic for subsequent articles.
Conclusion
I wanted to explore the topic of AI Agents for use cases specific to the needs of our clients at Etherion Consulting or even for internal use cases.
AI Agents can be specifically used for data validation and orchestration in a cost effective way. As I get my hands-on experience with AI Agents, I will expand on this article with a detailed tutorial and my experience with building AI Agents.
Exciting times ahead, especially for consulting firms like Etherion as we explore our options to add more value to our partnerships with clients on Data Engineering and Data Analytics challenges.
If you know more or are exploring more on the topic, I would like to hear in the comments to understand ways to make this topic more result oriented.
