Shawn Wang: evaluation

Transcript context

…I do think, I do think that you can build an agent that way. For me, I think you need to define the difference between. An MCP server plus client that is just a proxy versus an agent. I think there's a difference. And I think that difference might be in, um, you know, for example, using a sample loop to create a more richer experience to, uh, to, to have a model call tools while like inside that MCP server through these clients. I think then you have a, an actual like agent. Yeah. I do think it's very simple to build agents that way. Yeah. I think there are maybe a few paths here. Like it definitely feels like there's some relationship. Between MCP and agents. One possible version is like, maybe MCP is a great way to represent agents. Maybe there are some like, you know, features or specific things that are missing that would make the ergonomics of it better. And we should make that part of MCP. That's one possibility. Another is like, maybe MCP makes sense as kind of like a foundational communication layer for agents to like compose with other agents or something like that. Or there could be other possibilities entirely. Maybe MCP should specialize and narrowly focus on kind of the AI application side. And not as much on the agent side. I think it's a very live question and I think there are sort of trade-offs in every direction going back to the analogy of the God box. I think one thing that we have to be very careful about in designing a protocol and kind of curating or shepherding an ecosystem is like trying to do too much. I think it's, it's a very big, yeah, you know, you don't want a protocol that tries to do absolutely everything under the sun because then it'll be bad at everything too. And so I think the key question, which is still unresolved is like, to what degree are agents. Really? Really naturally fitting in to this existing model and paradigm or to what degree is it basically just like orthogonal? It should be something. I think once you enable two way and once you enable client server to be the same and delegation of work to another MCP server, it's definitely more agentic than not. But I appreciate that you keep in mind simplicity and not trying to solve every problem under the sun. Cool. I'm happy to move on there. I mean, I'm going to double click on a couple of things that I marked out because they coincide with things that we wanted to ask you. Anyway, so the first one is, it's just a simple, how many MCP things can one implementation support, you know, so this is the, the, the sort of wide versus deep question. And, and this, this is direct relevance to the nesting of MCPs that we just talked about in April, 2024, when, when Claude was launching one of its first contexts, the first million token context example, they said you can support 250 tools. And in a lot of cases, you can't do that. You know, so to me, that's wide in, in the sense that you, you don't have tools that call tools. You just have the model and a flat hierarchy of tools, but then obviously you have tool confusion. It's going to happen when the tools are adjacent, you call the wrong tool. You're going to get the bad results, right? Do you have a recommendation of like a maximum number of MCP servers that are enabled at any given time? I think be honest, like, I think there's not one answer to this because to some extent, it depends on the model that you're using. To some extent, it depends on like how well the tools are named and described for the model and stuff like that to avoid confusion. I mean, I think that the dream is certainly like you just furnish all this information to the LLM and it can make sense of everything. This, this kind of goes back to like the, the future we envision with MCP is like all this information is just brought to the model and it decides what to do with it. But today the reality or the practicalities might mean that like, yeah, maybe you, maybe in your client application, like the AI application, you do some fill in the blanks. Maybe you do some filtering over the tool set or like maybe you, you run like a faster, smaller LLM to like filter to what's most relevant and then only pass those tools to the bigger model. Or you could use an MCP server, which is a proxy to other MCP servers and does some filtering at that level or something like that. I think hundreds, as you referenced, is still a fairly safe bet, at least for Claude. I can't speak to the other models, but yeah, I don't know. I think over time we should just expect this to get better. So we're wary of like constraining anything and preventing that. Sort of long. Yeah, and obviously it highly, it highly depends on the overlap of the description, right? Like if you, if you have like very separate servers that do very separate things and the tools have very clear unique names, very clear, well-written descriptions, you know, your mileage might be more higher than if you have a GitLab and a GitHub server at the same time in your context. And, and then the overlap is quite significant because they look very similar to the model and confusion becomes easier. There's different considerations too. Depending on the AI application, if you're, if you're trying to build something very agentic, maybe you are trying to minimize the amount of times you need to go back to the user with a question or, you know, minimize the amount of like configurability in your interface or something. But if you're building other applications, you're building an IDE or you're building a chat application or whatever, like, I think it's totally reasonable to have affordances that allow the user to say like, at this moment, I want this feature set or at this different moment, I want this different feature set or something like that. And maybe not treat it as like always on. The full list always on all the time. Yeah.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.