Understanding AI Agents

January 17, 2025

Understanding AI Agents

Update in December 2025:

From Andrew Ng's Agentic AI course on DeepLearning.AI, the notion of AI Agents has been recast as "Agentic AI" which refers to LLMs being applied in an iterative or multi-step manner. In principle, one could use the same LLM in multiple sequential steps to refine its prior outputs, though there may be reasons in practice to use different models for different stages of generation. This reflects a generalization of LLMs as agents from the body of the original blog post here from the start of 2025, that described AI agents in the context of LLMs engaging each other in feedback loops. Indeed, consider the below example which shows 2 LLMs being used in sequence, one following the other to refine the initial outputs of the first via Reflection on the initial performance in order to produce a higher quality plot.

An idea to consider-- if one uses evaluation metrics to measure the performance gains associated with these iterative applications of LLMs, then one could imagine introducing a feedback loop into this framework so that N-iterations are applied until some performance criterion is appropriately met.

------------------------------------------------------------------------------------------------------

From January 2025:

The purpose of this post is to develop an understanding of AI Agents. They are a major development in deep learning for 2024. Conceptually, they leverage feedback loops between LLMs directing output from one into input to the other.

The below snippet from DeepLearning.AI's short course on AI Agentic Design Patterns with AutoGen is a nice example that demonstrates the idea with two LLM instances named "cathy" and "joe" who are comedians piggybacking off of each other's punchlines.

Below is a nice concrete instance of the termination criteria alluded to in the first abstract diagram for this same example. Once one of the LLMs produces output that includes "I gotta go" or "Goodbye" then the feedback loop between the agents ends.