In the middle of a browser QA testing process, the agent decided to wait 3000 seconds, then do some action, and then wait 5000 seconds. This makes me think there was a change in how agents decide the time they want to wait, i’ve never seen them wait for more than an hour for no reason. They probably wanted to wait milliseconds, not seconds
Steps to Reproduce
Go into a testing session for a feature for a frontend, that feature needs some time to load so the agent decides to wait. See if it mistakenly waits thousands of seconds instead of milliseconds
Hey, thanks for the report with the request ID and screenshot, that really helps.
I checked the session logs. There were no actual multi-minute blocking wait calls in this run. All tool calls (shell, mcp, delete) finished in 1 to 26 seconds. It looks like wait 3000 seconds and wait 5000 seconds were just the model narrating in its text output, not a real blocking sleep it actually ran. The shell sleep N takes seconds, and if the model had really called sleep 3000, that tool call would’ve hung for 50 minutes. We don’t see that in the logs.
Also, you’re on 2.5.17 (Feb 17 build), which is pretty old. @AndreasM1 mentioned in the thread they saw similar behavior on older builds and it went away on newer ones, so it’s worth updating to the current stable and checking if it still happens.
If on the latest version the model again decides to wait thousands of seconds, or it actually gets stuck in wait, send a new request ID and we’ll take a closer look. And if you want a quick safeguard right now, you can stop the agent and explicitly set wait limits in the prompt.
