Validating tool arguments before the call saved most of our agent bugs
Most of this team's agent failures were not reasoning problems but the model passing malformed arguments to tools. A strict schema check between the model and every tool — returning a short structured error so the model corrects itself — fixed more than any prompt tweak. Logging every rejection also reveals which tool descriptions the model consistently misunderstands, which is usually a documentation problem, not a model problem.