Tool-calling: would output schemas be useful?

[domenic]

At @michaelwasserman's prompting, I investigated whether in addition to inputSchema, any existing tool-calling APIs support declaring an output schema.

Looking at the MCP spec, they indeed have an outputSchema property.

Would this be helpful for our implementations? That is, is there any benefit to the language model, if the web developer tells it what format to expect the results in? It seems a bit unlikely to be a useful way to consume context tokens to me, but I'm not an expert.

Note that MCP is the only place I saw that included this functionality, so I suspect the answer is "it's probably not that useful" for a base language model interface.