What an agent actually is

Lesson 3.1 · 3 min read · by Vinay C

We'll cover the following

Overview
The loop
Tool calling mechanics
Memory
Summary

An LLM in a harness with tool calling is an agent, or very close to one. The question is who is in control.

	Who picks the next step
Workflow	Your code. "Retrieve, then summarize, then answer." The LLM fills in blanks.
Agent	The LLM, in a loop.

The loop

messages = [user question]
repeat:
    response = LLM(messages, tool definitions)
    if response has no tool call:
        return response            # done
    result = run the tool the LLM asked for
    messages.append(tool result)

Everything else (memory, planning, multi-agent) is built on this loop.

Tool calling mechanics

You send the LLM a description of each tool: name, what it does, and a JSON schema for the arguments.
The LLM does not run anything. It outputs "call search_orders with {id: 42}".
Your code runs the tool and feeds the result back as a message.

Memory

Short-term: the message history inside the loop.
Long-term: stored outside (a DB or vector store) and retrieved when needed.

Check yourselfWho executes the tool call: the LLM or your code?

Your code. The model only emits the tool name and arguments; your harness runs it, validates it, and returns the result.

Summary: say it in the interview

"An agent is a loop where the model chooses the next action. Tools are just functions my code runs on its behalf."