4 ms·
When I think about the structure, I automatically recall that when MCP calls are made, they do perform a search, but the model's tool descriptions and input/out
by jdw64 3mo ago
When I think about the structure, I automatically recall that when MCP calls are made, they do perform a search, but the model's tool descriptions and input/output are included together. So the list of tools becomes longer. Reflection might reduce the cost of writing wrappers, but it seems likely to use more tokens. There are also security concerns, and there are MCP host issues as well. I don't know exactly how the agent calls tools, and there's caching involved.
The reason everyone is using similar techniques for these tools is probably that the harness environment surrounding agents differs from company to company, making it hard to specialize. It's good to come up with something new, but ultimately, industry standards are what matter
- ddddazed 3mo agoThanks! It shows me that you really understood what I was trying to say. And yes, the issue of “security” is a concern I share in fact, it involves memory injection, and more than the cost on the JSON side, it’s this aspect that doesn’t quite convince me. I already have everything up and running (as an experiment), and the possibilities are expanding. Your second valid point is about standards, but here I’m sticking to my simple experiment, and I was interested in hearing perspectives like this—not to propose a new stack, but rather to explore whether a new layer could unlock new opportunities. Legacy software or the ability to modify graphics at runtime come to mind, this part has also been tested. It works, but I’ve never seen a real-world application for it, and it’s not my job to create products or identify market needs. The interesting part and the one I was seeking feedback and ideas on—was the technical aspect. Thank you so much for your response.