The SOTA models have excellent instruction following capability and the ability to output in any format you want including JSON. That's all you need from the m

The SOTA models have excellent instruction following capability and the ability to output in any format you want including JSON. That's all you need from the model to be able to use it in an agent. Tell it to output commands in a given JSON format. I assume that Mistral's API already allowed you to define the system prompt, right?