Call My "Agent"
If you’ve been entangled in the LLM whirlwind over the past few years and have started to use the word “agent” without a firm understanding of the concpets this word attempts to capture (much like the author of this post), the follwoing paragraphs hope to unpack the relevant details.
Working Model
We will take the above model for the LLM interface as a given (to delve into this black box, see Internals of a Large Language Model (LLM)). This interface can be enhanced by augmenting capabilities (“Tools”) to inspect and mutate content. One mechanism for the LLM to access these capabilities is via the Model Context Protocol (MCP).
The Loop
Chaining several of these calls together, with a verification step that decides whether to iterate or stop, is, at a high level, an “agentic” loop. Press the check to keep iterating, or the cross to fail out and restart.