Sometimes it happens that the agent tool usage gets into an infinite loop. Even as I have max_iters capped at 5. This infinite loop can cause an insane high token usage.
So is there any setting that can limit the total token usage of the crew? Like the crew shuts down when this limit is reached.
This could save a lot of unnecessary costs.
@Raja_Speet Can you confirm that the loop is not caused by agents delegating the work between each other? It’s hard to say without seeing the chain of taught of your crew, but the problem might be that an agent is iterated 5 times, but then it delegates the work to another agent, which is also not able to get the final answer, and this is why you see an infinite loop. If this is the case, set allow_delegation=False to all agents.
Adding to above In my experience it was primarily due to the incorrect input structure and the tool output structure. In the previous version 0.51 it was difficult to figure out. In the latest version crew gave this error. So generally if any error happens Just anticipate multiple exectutions from crew
I can confirm there was no delegating involved. The process I use is also sequential.
I get this problem sporadic. Got it when I used a custom tool, but also most recent when using the standard FileReadTool. After I got the problem and shut down the execution manually and I ran it again, I got a correct process.
As tool output the agent returns: I tried reusing the same input, I must stop using this action input. I’ll try something else instead.
And afterwards starts over with calling the tool again, and again…
@Raja_Speet What about setting an instruction to an agent or task something along the lines of: If the tool doesn’t provide a final answer after 3 attempts, stop using it and proceed without it.
Thanks, I will try that!
what model are you using for this. many open source models are terrible at function calling. I always test with something like gpt4omini when that sort of thing happens with open source models.
The parameter is max_iter not max_iters.
also you can try max_retry_limit which is for errors.
I also use gpt4omini and max_iters was a typo indeed.
max_retry_limit is standard at 2, right?
@Raja_Speet Yes, the max_retry_limit parameter is set to 2 by default. See the docs.