The agent mental model

The intellectual evolution from chat to agent

Let us briefly review the history of how LLM technology evolved over time.

LLMs begin as “text generators”

Remember that early modern LLM use was essentially:

Prompt → LLM → Response

The model could explain, summarize, draft, and answer questions, but it could not independently interact with the outside world.

Reasoning improves

Researchers found that models often performed better on complex problems when they decomposed the problem into intermediate reasoning steps. Conceptually:

Question → Think → Think → Answer

But reasoning alone has a limitation: the model is still reasoning only from what it already knows. It cannot check a database, search the web, send an email, or inspect a live system.

Give the LLM tools

The next important idea was to let the model call external functions:

LLM → search()
LLM → calculator()
LLM → database()

Now the model can act on the world, or at least retrieve information from it. But this raises a new question: How does the model decide when to think versus when to use a tool?

ReAct: Reason + Act

In 2022, researchers from Princeton University and Google Research provided a very influential answer: interleave reasoning and actions rather than treating them separately. The basic loop looks like (ReAct):

Reason → Act → Observe → Reason → Act → Observe → … → Answer

Goal: Find the cheapest flight that satisfies my travel constraints.

Reason: I need to know the dates and available flights.
Act: Search flights.
Observe: Here are the available flights.
Reason: Two options satisfy the constraints; I should compare baggage fees.
Act: Look up baggage policies.
Observe: Airline A charges $70.
Reason: Airline B is cheaper overall.
Answer: Recommend Airline B.

ReAct becomes the basic architecture of an agent

Once you have this loop, you are very close to what we now call an agent:

Goal
  ↓
LLM decides next action
  ↓
Use tool
  ↓
Observe result
  ↓
Decide what to do next
  ↓
repeat until goal is achieved

Modern agent systems add things around this basic loop—more sophisticated planning, memory, permissions, guardrails, multiple tools, and sometimes multiple agents. OpenAI describes agents as systems in which a model operates with instructions, tools, and guardrails to execute workflows.

The diagram below lays out the evolution:

A comparison of chat versus agent

Compressed into one comparison, the shift looks like this:

Exercise: Chat or agent?

For each task below, decide whether it describes a plain chat interaction or an agent interaction.

Check your work: rows 1 and 3 need nothing beyond the text in the message, so plain chat is enough. Rows 2 and 4 require opening files and verifying a result against them — that requires an agent.

Once you can tell the two apart, the more useful question is what changes about your own job once an agent is doing the acting. The next page takes that up.