The Loop

What changed

A chatbot receives your message and produces text. One turn, one output, and nothing happened in the world.

An agent is given a goal and then runs a cycle:

1. Decide what to do next
2. Call a tool
3. Read what came back
4. Decide again

Repeating until it believes it is done or it runs out of budget.

The underlying model is exactly the same. What changed is that its output now triggers actions, and the result of those actions comes back as new input.

Tools are the whole difference

Without tools, a model can only produce text. It can tell you what it would do.

Give it tools and it can search the web, run code, read and write files, query a database, send an email, or book a meeting.

That is the entire step change. Not intelligence. Reach.

And reach is a multiplier on every failure mode you already know about. A model that fabricates a citation in a chat window wastes your time. A model that fabricates a customer record while writing to a database has done something you have to undo.

Read the trace

Good agent tools show you the steps. Learn to read them, because the trace is where the diagnosis lives.

What to look for:

• Where did it go wrong first? Everything after inherits it
• Did it verify anything, or accept the first result it found?
• Did it loop, trying variations of a failing approach?
• Where did it decide it was done, and was that condition right?

That first question is the important one, because failures propagate forward. The visible error at step twenty is usually a symptom of a quiet wrong turn at step three.

Completion is a decision it makes

Worth pausing on: the agent decides when it is finished.

It stops when it judges the goal met. That judgement is another prediction, and it can be wrong in both directions. It can stop early, having answered a simpler question than you asked. It can grind on past the point of usefulness.

So "it finished" is not the same as "it succeeded". A completed run is a claim to check, not a result to accept.