Agent can get caught in a tailspin where it’s changing code, running tests, change code back ad infinitum.
Creating a new conversation, having it reflect on its failures can sometimes solve the issue.
Agent should be able to detect when it’s chasing its tail, and try at least once a new conversation to recollect itself.
It shouldn’t be too hard to detect these situations from cyclic code changes and the looping self commentary. It’s like a roomba caught under a chair - you can hear from the next room it’s in trouble.
Thanks for the great suggestion! You’re right that detecting when Agent gets stuck in these loops would be super helpful - like a roomba escape mechanism
While we don’t have this feature yet, I’ll definitely pass this idea along to the team. The current 25 tool call limit (mentioned in our Agent docs) helps prevent infinite loops, but having smarter detection of cyclic behavior would be a nice improvement
In the meantime, starting a new conversation when you notice Agent going in circles is a good workaround
