MCP and the Dawn of Agentic Intelligence | Gabriela de Queiroz & Shub Argha, AI By the Bay 2025
Hi folks, I'm Schu. I am the head of solutions engineering at arcade.dev. I work with all of our enterprises and deploying agentic applications. And I'm really really excited to talk about context with you all today. And I'm joined with >> my name is Gabriela. I am the founder of F-02 Labs where I work with startups on the AI and developer relation strategy. I used to work at IBM on the open source side and then Microsoft. Very excited for our talk today
So let's get this going. >> So put it all in context. Context matters. >> Intelligence comes from context. We become intelligent because we are surrounded by our memories, our tools, culture, stories, feedback and environments. For us, context is our whole world. Intelligence is a dance between a mind and its world. AI doesn't have that same luxury
And without a world to behold, the mind grows weak. So there is no context. So AI needs the model context protocol or MCP. MCP is an open protocol that standardizes interface between AI systems and tool providers. Now, I put standardizes in quotes not to piss people off, but because sometimes it pisses me off and because MCP is growing so fast and it's a growing protocol and as we're building so fast, some of these pro uh some of these standards aren't being followed and they're being followed differently by different people. But overall, so overall it defines a few things. So some responsibilities around server client uh resource fetching the two invocation and the capability negotiation. >> And this is what a standard MCP client and server architecture looks like
On the left you've got clients like your chat interfaces or your chat agents like chat GBT. You've got your coding agents like cursor or cloud code as well as any other agentic application or AI agent that you might make. And then on the right you've got your server. So this is all of the things that you might need access to. That's your data and file systems, your dev tools, as well as any of the productivity tools that you all might use. Both the client and the server have to meet in the middle with MCP and they negotiate their capabilities, what they can each do. First, the MCP servers capabilities. very quickly
MCP servers may optionally implement tools, resources, prompts, and other logging facilities. And the same goes for the client. It may support roots, sampling, elicitation, experimental stations. So in summary what we are talking about here so far is that the client has capabilities the server has cap has capabilities as well and they negotiate but you know at the end of the day sh all right Gabby this is all kind of boring we came here to talk about context and that context is all that matters So, not to mock you all, but here's the problem. Your AI agent can't do a single push-up. This is your AI agent today. The LLM that's powering it has watched thousands and thousands of hours of 80s workout videos. It can describe to you every muscle, every routine, perfect form, but they can't actually make the gains at the gym
And we've built these extraordinary models like this. They can write workout plans, make detailed meal plans, keep track of your progress, but they don't know your world. They have never been to the gym. They have no context. And because of that, your AI agent is weak. All right, Shu, got you. Uh, you're telling me that the AI agents know things, but they can't do things, right? So, they cannot access your whole world. Uh, they cannot do, you know, your loved berries class
>> Exactly. >> Exactly. You're right. And that's why your AI agent needs a personal trainer. And that is MCP. It's that legendary trainer who looks at your AI agent and yells, "Kid, you need a program. A program that finally lets your AI agent work out exactly in the way that it describes in such intense detail. Now, we're ready to train." >> All right
All right. So, this is the MCP planning or training plan. Same things that we talked before. We have the servers, we have the clients, we have capabilities, but you know, let's go ahead and get into the exercises and our workout plan. >> Let's do it. >> So, exercise one, tools. Tools are the biceps of your agent. This is how AI actually reaches into the world and does stuff
Send emails, for example, book flights, write to databases, flex those API sh Exercise two, resources. Resources are your AI's eyes. You can give it your documents, your calendar, your codebase, and then suddenly it'll be planking over your entire knowledge graph. Then we go to exercise number three, which is prompts. Pretty much everybody here is very familiar with prompts. And prompts teach you or teach your AI the form. This is how you tell it. You are my travel planner today
Or maybe you say something like, you are my email whisperer. Stretch out those subject lines. Um, prompt bend your agent into the perfect roll. Okay, so we've covered all of the exercises for the server. Now, let's move on to the client exercises. The server's exercises seem pretty simple. You've got the hang of it without breaking a sweat. Now, let's you put your AI agent through the berries burner workouts that Gabby described with that exercise four sampling
I'm really excited about this one actually. Sampling. This is your AI agent's inner voice. This is what it's doing when it zones out at the gym and is staring into the wall into the corner after an exercise. Sampling is your agent's ability to choose, to reflect, and decide. It allows the server to use your client's LLM to make decisions. This is where your agent whispers to itself, join the devolution. >> Then we move to exercise five, which is elicitation
This is your AI agent saying, hey, can I get a spot? When your agent doesn't know something, what does it do? It ask for help. Right. >> All right, Gabby, let's take a quick off break. Elicitation is also how your AI agent asks for access and off. If it needs to access something that it it is not authenticated or authorized to do, it will stop and elicit from you access. And it does this by asking you to click on a URL and log in just like you've done many other times before. So you can use the tools at the server. Arcade built this into the MCP spec
It's already been adopted by many MCP clients like VS Code. Uh, and it will be part of the official release next week. Today's the 18th. So, literally seven days from now on the 25th when the new spec comes out, it'll be it'll be in there for you. >> All right. So, let's move on to the exercise number six, which is roots. Roots keep your AI in this workout zone, right? So if you are in if it's cardio time, you stay in the cardio area and remember the boundaries the folder this folder only or this workspace only or don't go wandering into your slashetc that's where my friends injuries can help can happen. >> All right now let's do a bonus exercise tool chaining
With tool chaining, this is where your AI has every has trained every muscle of context at the server and at the client. It it then performs the ultimate routine tool chaining. Multiple steps, multiple tools. It can use multiple resources and prompts, sampling, elicitation, all of it firing, all within the MCP protocol. This gives you the ability to build AI workflows without having to add additional application code to your agent. This is the full aerobics finale. >> So ladies and gentlemen, this your agent now [laughter] it's shredded on context. It's oiled up with capabilities and it's absolutely flexing across your entire AI workflow
>> So, let's all remember bigger models will not save you. More GPUs will not save you. Sweat bands and tight masters definitely won't save you. >> Only one thing gives AI agents true power. Context. That's all that matters in AI agents. >> And with Arcade, Arcade.dev, you can put your AI agent through the whole MCP program in only seven minutes. >> Arcade is an MCP runtime and gateway
You can build secure authenticated MCP servers with Arcade secure MCP framework. >> And you can deploy and run these servers with proper security and governance. And with Arcade, you can enable anyone across your organization to use MCP servers. Multi-user ready out of the box and ready to be used with any AI agent or MCP client with proper authentication, authorization, security, and governance. >> Build Jack secure AI agents at rk.dev. >> Thank you. >> Thank you folks. [applause] >> That was fun
All right, I think we have some time for questions. >> That's correct. You can have um spend like maybe like five minutes of u yeah questions and answers. So, anybody has any question for Gabriella and Shup? Yeah, no. Quiet >> questions. Um I can just like give you some uh the behind the scenes why we are using this whole workout analogy here. Uh do you want to share a little bit of like what is coming up? >> Lots of things are coming up. Uh but we're really excited about the launch of the next MCP spec that's coming in a week here
Uh it's going to include a lot of new features and capabilities that will allow you to take even more secure uh build more secure MCP servers and uh use them directly into your AI agents. Um and we're of course we just launched our MCP framework directly on Product Hunt. So we're really excited about that and everything that's coming with it.