What is an AI Agent: Understanding Agentic Systems

One of the most debated questions in recent times is: What is an agent? If you use ChatGPT, ask a question, and get a response, does that mean you are interacting with an agent?

Who you ask this question to can get you very different answers, and most of them will be coming from the implementation angle. None of these definitions are wrong, and each perspective has a valid argument.

As a general purpose definition, an Agent is a system that takes input from a human or another system and acts autonomously to achieve an outcome.

As Andrew Ng, says in this article, rather than having to choose a system as agent or not in a binary way, we all can agree that there are different degrees to which a system can be agentic. It is better to focus on the agentic design patterns and characteristics.

Key Components of an AI Agent

With the GenAI explosion in recent times, we have LLM as the center of any agent, and it is practical. LLM handles the most critical Planning part of the Agentic system.

Agentic system typically includes below components,

Planning: An agent needs a cognitive component, typically an LLM, that enables reflection, reasoning, self-refinement and decision-making.

Memory: The ability to store and recall previous interactions is crucial for contextual awareness and continuity. This is served typically by an external database

Tools: External systems agents uses to perform action like retrieve data, perform calculations or take actions

Execution Framework: Orchestrates the flow of tasks, calling the LLM for decision making. It integrates feedback for self-improvement and enables multi-agent co-ordination.

Agentic vs Non-agentic

Not every system using a Large Language Model (LLM) qualifies as an agent. Anthropic differentiates it as Workflow vs Agent. I prefer Andrew Ng’s version of calling it as Agentic workflows vs Non-agentic workflows.

• Non-agentic Workflows: A system that follows a pre-defined path. The LLM plays a specific role in this flow, but it does not determine the overall direction of the flow.
• Agentic Workflows: A system where the flow is not predetermined. Based on inputs and conditions, the LLM can decide which path to take next. An example is a loop that continues till the desired outcome is achieved. Essentially, an agent “thinks” and determines the next step to achieve the desired outcome.

Final Thoughts

At the end of the day, labels and agent qualifications don’t matter as much as the outcome. What’s most important is that these systems enhance productivity and deliver meaningful results. However, understanding the distinction helps us design better, more efficient AI-powered solutions.