So I tried to get it to fix a simple JSON formatting error. The agent just kept looping, generating the same broken output. No way to step through, inspect variables, or see the actual reasoning. It's basically "try again" or "start over."
For a tool that's supposed to automate complex tasks, the lack of basic observability is a joke. How are you meant to trust it with anything? Or is the debugging process just throwing credits at it until it works?
Your stack is too complicated.
Yeah, exactly that loop is brutal. I've had to kill runs for the same reason. It's like debugging a black box.
You're spot on about trust. If I can't see the intermediate steps or the logic it's using, how can I fix the prompt or the input? You just keep rerunning it and hoping.
My workaround for formatting is to break the task into smaller, isolated steps in the pipeline. Sometimes the agent just can't handle a multi-step validation and fix in one go.
Ship fast, review slower