AI Agents Can Get Stuck in a Loop

AI agents can get stuck repeating the same step over and over. Always set a max step limit so they don't run forever.

Some examples

• AI tries the same broken step 50 times.
• AI keeps searching the web with no answer.
• AI may rack up costs while looping.
• A step limit forces AI to stop and report.

Try it!

Watch an AI agent. If it does the same thing 3+ times with no progress, stop it manually.

Recognizing and stopping a runaway agent

A runaway agent is one that keeps doing the same failing action over and over — like a robot vacuum that keeps bumping into the same wall. It's frustrating, wasteful, and sometimes dangerous if the repeated action has real-world consequences (like sending the same email 50 times). Runaway loops happen when the agent's step limit is too high or missing entirely, when the agent doesn't check if it's actually making progress, and when there's no human watching to intervene. Spotting a runaway is usually easy: look for repeated identical actions in the log, a rapidly climbing cost meter, or a task that's been 'in progress' for far longer than expected. Stopping one is also easy if you've designed in a kill switch — a manual override that cancels the agent immediately. The real lesson here is that kill switches and step limits aren't emergency features. They're standard features that every agent deployment should have from day one.

• Cap the step count: no agent should take more than N steps without a human checkpoint
• Check for progress: if the last 3 outputs look identical, flag and pause
• Build a kill switch: always have a way to stop the agent instantly from the outside
• Review the log: repeated identical actions in the log are the clearest runaway signal