Hello everyone. I'm Ganesh Swami. I'm one of the co-founders of Covalent. And today I'm going to do a quick walk-through the AI agent SDK. This is a TypeScript SDK that we've built for developers and you can use this SDK to launch agents.
So these agents are can perform any kind of task, but it's not just these agents. You can also make these agents autonomous and we're introducing this concept known as a zero employee enterprise, ZEE in short. And what this does is that it brings together this swarm of agents that can iteratively and collaboratively collaboratively solve any kind of complex problem. And you can see all of the inner workings of of of how these agents are able to collaborate. And so this is relying on a technique called chain of thought models.
This is a new way of using these LLMs. So what you're doing is you just like if you were to to ask someone give me a solution to a math problem, they're not going to give you like a Google search result, right? Google search results you type in a query, gives you an instantaneous result. Here what it does is they probably go away and then take a couple of minutes, couple of hours and then work through a couple of solutions, think through the problem, reason through the problem and then suggest a solution. And if it's not able to provide a solution it's able to show you the inner workings of how far it was able to get.
So that's basically what's happening with this new form of LLMs. And so we rely on this technique to have these agents basically ping-pong and challenge each other and bring autonomy to their inner workings. And this is what makes them super exciting. So uh jump right into the code. So, the first thing I would ask you to do is this agent SDK is fresh off the press.
Uh we literally just released this as open source today. And so, the first thing I would recommend you read is the blog post that announces the ZEE concept. And so, I'm going to share my screen now and then the rest of the this recording will have with screen share. And so, this blog post is just on the covalent h2.com.
So, you can just go to the blog and see this is the agent SDK version two. Uh you know, we've had to rebuild this SDK from the ground up because a lot of the assumptions we've made during during the like the first version of the SDK, which by the way was only in December and crypto AI agents, all this stuff moves like so fast. So, you have to rethink and you know, challenge your assumptions at all points. So, this is basically telling you what's going on here. We've doubled down on the crypto use cases as opposed to building a generalized you know, AI agent tool.
And so, it just tells you how you can do all of this stuff. And then the next thing what you need to do is you need to go to the to the GitHub repo. So, this is uh right here. Actually, before going to the GitHub repo, I'll recommend you go to the docs, which gives you an overview. So, go to cxt.
build and this is the overview of how the SDK works, the different components, how do you actually build an agent, how do you piece together tools, and how do you build a ZEE workflow. So, let's walk through some of the concepts here. So, the first concept is an LLM, right? So, what an LLM does is it's just a thin wrapper over one of the language models. But, the beautiful thing here is you're able to access about a dozen different LLMs.
So, what you see here is the OpenAI LLM, Deep See, Grok, Gemini, and then they obviously take different environment variables. And so, you literally just have to initialize an LLM with the provider and with the the name of the actual model. So, that's what's pretty interesting about this. The second thing you need to second concept you need to be aware of is what is known as an agent. And so, an agent it basically has a name.
It obviously uses a model. It has a description of what the agent is supposed to do and a set of instructions on, you know, what is the what is the basis of this agent. And so, here this specific agent this example is a reporting agent. It relies on OpenAI's 4.0 mini and this agent is responsible for generating reports and it generates a report on the current state of the company.
So, that's basically how an agent does. Now, what's also special is that these agents are completely stateless, which means they don't have any kind of memory or something. And so, here or they cannot really access the outside world. And so, here this is where you would add tools. So, you would just go add some tools.
We have one tool currently built, but there's a lot more being built behind the scenes and you're also welcome to contribute tools to the the SDK. So, one of the tools we have is the Gold Rush on-chain data tool. And so, this allows you to get token balances, NFT holdings, transaction history, all kinds of things. And so, when you go back to the agent and you give it the tools, then what the agent is able to do when you run the agent using agent.run, it's able to try it's able to solve the problem and the instructions that you've given, but also use these external tools.
So, it's important to describe these tools in a very you know, in a very descriptive way so that the LLM knows, "Okay, at this point, I have access to this tool and therefore I can basically outsource the thinking this tool." And then the framework automatically calls this this basically function calling and so that's like the cool thing about this this the agent stuff. And here this is some description on you know, what the tools can do. Not only can it call external APIs like the Gold Rush API. The Gold Rush API gives you access to about 100 different blockchains.
It's also another covalent product. But the other beautiful thing is that it can also structure the output from these these these agents. So, this how you create a tool, you create a schema and so on. And then the most exciting thing is actually the ZEE workflow. So, what the ZEE workflow does is you combine the set of agents and then you give it some kind of like goal and then what essentially this workflow does is internally it has a router that basically does ping pong between the different agents and their respective tools.
And any kind of memory that is stored is stored on the ZEE layer. And so, in theory what you can do is you can actually use Gemini for let's say a multi-model agent and then you can use let's say Deep Seek because it's really good at certain set of problems. And you can actually get these two models actually speak to each other and ping pong you know, and iterate and then solve this problem. And so that's really what's cool about this whole thing. And then what I I jump to the actual repo and give you a layer of the land.
All of this is open source, so what we are looking is to get builders to either build using SDK or directly contribute to the SDK. So I'll get into the layer of the of the repo in a bit. So here what we have is this this thing is completely open. So just go to covalent hgr.com AI agent SK.
It's also linked from the blog post. It's also linked from cxt.build. It should be everywhere. It's from the covalent homepage.
It's hard to miss. You know, if if you miss if you're if you still miss it, I think we've done a bad job here. But you know, it's it's available. It's easily accessible and then you can outsource this so you can you can you can check out this this code and then build it yourself and so on. So what this does is basically you can go to the the packages and within the packages you have the AI agent SDK.
And then you can look at the source and then you have all of these LLMs. You have the base and the state. So let's take for example, you want to contribute a new LLM. So go to the LLM and then you open up index.ts.
And so what you see here is you have some definitions for OpenAI's models, Deep Seek models, Grok model, Gemini model. And then you know, it's missing Anthropic. Anthropic is another fantastic model. It's missing other kinds of like hosted models. Maybe even So another nice contribution could be is referred to Akash's inference API or one of the other providers out there.
Llama is another really cool kind of model. And so that is something you can contribute to. And then something else you can contribute to is you go back to the tools and then you have the list of tools, right? So, here we have the Gold Rush tool. So, for NFT balances, token balances, and transactions, and so on.
But, you can add any other kind of tool, right? Let's say you want to add a tool to do contract signing. Or, you want to check the a tool to maybe verify if the particular transaction is safe. All kinds of things, right? So, that's something definitely we're looking to build a big library of tools.
And then there's a bunch of other stuff. For example, there's the actual like workflow in the ZE folder. So, that's basically the workflow how it interacts with the problems. And then you have the agent itself. So, this is basically how like all of the materials inside inside the agent including the routers and so on.
So, that's basically what's going on with the the repo. The documentation is also available here. So, you go to the docs and you can also make documentation contribute to the documentation. Though, that's more like I don't know if that's worthy enough of a contribution. But, definitely if you find any kind of gotchas, you know, open up a PR and it's much much appreciated.
There's two two more things I want to talk about, which is how do you actually use this library and what are the use cases you can build using the API using the SDK. So, first thing is like how do you actually use this? So, to use this we've built a very neat kind of tool called um create Z app. So, you can just go to the quick start and do NPX CovalentHQ/create-z-app. And what this does is it asks you about a bunch of questions.
Right now, there's one template, but we'll have soon have multiple templates, and this creates a fully bootstrapped tool uh to create a ZE. So, you know, it just gets everything out of the box. Now, this is not very different from create React app or create Next app if you're familiar with those kinds of like um you know, helper helper tools. So, that's basically what's happening here, and then you get the whole source code, and then you can like, you know, use the agent and so on. So, this is like something that's really really cool.
And the final thing I want to talk about is what kind of contributions are we uh looking at. So, what we have here is something known as ZE use cases. It's on the cxt.build uh documentation site. And you go here, uh there's a whole bunch of use cases that we've uh written uh on what we would love to be love to see built with the SDK.
And so, something like this is is meteor. Something like this is, you know, uh worthy of a uh of a you know, a grant and uh further, you know, investment. And so, uh this is going to be almost not just using SDK, but also contributing to the SDK itself in terms of the tooling, in terms of the logic, in terms of like debugging. Uh maybe there's some specialized LLM that you require. Uh maybe it's going to create a chart.
So, you need some kind of like uh you know, uh multimodal. Maybe it's uh uh maybe it's like like notebook LLM or something that that creates a podcast after it, you know, it gives you a voice message. Uh all kinds of things. So, I think there's no limit to what you can build with this. So, this is what makes it super super exciting.
So, good luck, hackers. Uh I just want to leave you with one final thought. Uh I started Covalent uh at a hackathon. So, you know, there's one one weekend, I was just able to sit down and uh build something in 2 days, uh which was incredible. And uh Covalent has come a long way since that that uh one weekend.
And I really wish that uh for all of your projects to uh grow to this kind of scale. So, just telling you that just because it seems like a hackathon project doesn't mean that it's a throwaway. It could be the start of something it could be you know that Kindle that lights up a big fire. So, I wish you all the very best and happy hacking. Thank you.
Automatic transcript — names and jargon may be misspelled.