Loop

What is a loop?

You know the feeling: you ask an AI to tweak a sentence, it changes it, you don't like it, you ask again, and it changes it back. An agent "loop" is exactly that — it's trapped in act-check-not-good-enough-act-again, repeating the same kind of move while never reaching the goal.

Why are loops dangerous?

They burn time and money
Every turn spends tokens and API calls. An infinite loop is an infinite bill.
They're busywork
The agent may rewrite the same code or click the same button again and again, with the problem not one bit closer to solved.
They drag the whole system down
When several agents each get stuck looping, they eat resources and jam up everyone else's tasks.

How is a loop different from a deadlock?

Deadlock is "nobody can move"; a loop is "moving constantly but going nowhere". One freezes, the other spins — opposite symptoms, both nasty failures.

How do you prevent it?

Cap the iterations
Give each task a max step count; when it's hit, force a stop and report back.
Detect repetition
Log the last few actions and results, and if it's spinning in place, trigger a breaker or bring in a human.
Fix the goal
Split a fuzzy objective into small, verifiable steps where each one can be marked "done" or not.

Bottom line: a loop is an agent running in circles — fast, but never reaching the finish line.