New Ethereum talks, every Monday. The week's conference uploads by event, in your inbox.

Loading player…

Ali Serag & Bohao Tang I Build your First AI Agent With Flow Foundation I Agentic Ethereum 2025

ETHGlobalThu, Oct 9, 2025, 12:00 AM

Join Ali Serag and Bohao Tang from the Flow Foundation for a hands-on deep dive into building your first AI Agent on the Flow blockchain. Learn why Flow’s unique features—including resource-oriented programming, atomic transactions, multisig account management, and capabilities—make it an ideal chain for sophisticated AI-driven dApps. Discover how AI Agents can hold wallets, communicate, and bundle transactions while securely granting granular permissions. Follow along with a practical demo using a lightweight fork of Eliza integrated with Flow. If you’re ready to push the boundaries of Web3 with AI autonomy, this workshop is a must-watch! 00:00 Intro & Agenda 00:47 What Are AI Agents? 01:26 Agent Components & Personality 02:33 Why Build AI Agents on Flow 03:41 Atomic Transactions & Bundling 04:14 Multisig & Sponsored Transactions 05:30 Flow Account Management Explained 06:28 Granular Capabilities & Permissions 07:30 Lightweight Eliza Fork Overview 09:37 Demo Setup & Repository 11:17 Environment Setup & Config 13:12 Run the Agent & Debug Mode 15:51 Agent Wallet vs. User Wallet 17:35 Flow Actions & Plugin Examples 19:23 Batched Transactions & Scripts 20:10 Practical Tips & Best Practices 21:06 Where to Ask Questions & Get Help 29:37 Pitch Training & Product-Market Fit 30:18 Good Luck Building! _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 🤖 *Agentic Ethereum* Agentic Ethereum is an online, async hackathon that runs from Jan. 31 to Feb. 12, 2025. With over $175,000 USD in prizes, Agentic Ethereum is a great way to dive into the world of autonomous AI agents and explore how they can revolutionize projects on Ethereum. Start the year with groundbreaking innovation! Watch the full playlist: https://youtube.com/playlist?list=PLXzKMXK2aHh7jxP_flgRf5nZ_WDgFpBy3&si=RZozbd9C5KrG6Bf3 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ✅ *Follow Ali Serag* X: https://x.com/AliSerag0 ✅ *Follow Flow Foundation* X: https://x.com/flow_blockchain ✅ *Follow ETHGlobal* X: https://x.com/ETHGlobal​ Warpcast: https://warpcast.com/ethglobal Website: https://ethglobal.com YouTube: https://www.youtube.com/c/ETHGlobal _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _Are you interested in Ethereum development and entrepreneurship?_ 👉 Sign up for the next ETHGlobal event: https://ethglobal.com/events 🎁 Get exclusive access and perks with ETHGlobal Plus! https://ethglobal.com/plus 📣 Want us to throw an event in your city? Tell us where! https://ethglobal.com/city _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Transcript

Okay. Hello everyone. I am very happy to see you here happy to see you here . We have something very exciting . We're going to look at AI agents, one of the hottest topics in the Web3 community right now, how to implement them in Flow, and how to launch your first agent .

Today, I am with you, who leads the Devrel division for the Flow Foundation in North America , and my colleague Bohao Tang, who leads the Devrel division for the Asia-Pacific region . We start our agenda by explaining what these AI agents are . Why is Flow considered the best blockchain blockchain , if not the best, for building AI agents ? After that, Bohao will provide a demo of how you can get the best features to build complex agents . Right?

Agents are going to change everything . In recent months we have seen an explosion of interest, from the creation of Eliza, to AI 16Z, to billion-dollar AI coins, to Google writing white papers. But what are they really ? Why are they important? AI agents are autonomous systems that can provide services.

They can provide advice. They can do the work. It is said that these will replace all services such as writing essays, helping with grant applications, or getting personalized medical advice, with specialized human agents like Mechanical Turk, Fiverr, and Apork . In fact, a bill is currently being introduced to give AI the ability to issue prescriptions . It's unclear whether it will be accomplished , but they are poised to transform large industries and verticals.

They are not only responsive, but also efficient. They are constantly learning learning and adapting, and are available 24/7. This is one of the things that excites me the most right now, there are four key components that make up an AI agent. I'll make things a little easier for you . You can find these four components on sites like Eliza, an open source platform that helps anyone build their own AI agent.

Well, what does Elisa actually offer? This is a platform that allows you to create an AI agent with brain-like functionality . When I say brain, I mean you can connect an LLM using Deep Seek or OpenAI's ChatGPT or the Anthropic API or Claude . This allows it to receive and react to the tokens it is sent and learn how to understand the world , right? It uses those different AI models .

Like a child, you can shape your personality, and Bo House is going to give a small example of this to give a small example of this , in which you can define yourself as this is you . This is where you came from, son. This means that if someone asks you this in this context, you should respond this way. By teaching it how to act in different settings and contexts, it will easily adapt to new situations . Yes, you can teach it what style it should communicate in .

You can make it more elaborate, or artistic, or even tricky. The second characteristic of these AI agents is that they have mouths and ears to communicate with humans. This happens by integrating APIs of common social apps like Twitter, Telegram, WhatsApp, etc. They have hands, which means they can perform tasks beyond their usual tasks. They can check your Google Calendar , check price changes , schedule appointments, search online, and more .

They can analyze your Twitter profile to understand your communication style and act accordingly. They can look at GitHub commits, make PR requests , and work in many areas like this. Lastly, they have a wallet, which means they can send, receive, invest, and manage funds for you. By combining all of this, you can create great companies and characters with unique personalities who can perform various tasks . In short , they have the potential to become new employees or consumer agents.

So, why is Flow a great place to build agents? Flow has a unique transaction model that allows multiple transactions to be chained together . If any transaction in that batch fails , the entire transaction will be canceled. It allows atomic transactions . When it creates a transaction , it divides its responsibilities into three: proposer, payer, and authorizer .

Sorry, approver approver , proposer, and payer. So, at the transaction level, if you want to do sponsored transactions to create experiences without case fees, you don't need to have a complicated Paymaster contract or something like ERC-4337 . You can do it directly . Things like package transactions are really exciting because imagine you have an AI agent in a video game and you trade or barter with it , the World of Warcraft border menu appears. If you want to exchange five shields for three gems , you do n't want to have to make a separate transaction for each on-chain asset being exchanged.

You'd want to do them all at once want to do them all at once , wouldn't you? There are many more examples of what you can do like this. There are also multi-sig transactions , which give these agents complete autonomy . There are many situations where you need to have complete security and control over what your agent says. For example example , imagine you are in a Telegram chat , and you and your friends are investing in various meme coins together with an AI agent .

Maybe you want that AI agent to recommend different meme coins and if 70% of people say, "Okay, we approve this transaction," then that's what you want it to happen . It will come true eventually. Therefore, if 70% of people sign the proposed transaction , it will be executed be executed . Maybe you and your partner manage the funds together , or you have an AI agent that makes high- risk trades ; You may want it to check with you before trading more than 20% of your asset value. So, you sign that multi-signature transaction , and it goes through.

This allows for a lot of customization . Another great feature is account linking, also known as account management . This is very unique to Flow , which you can integrate with AI agents . So, account management essentially allows for joint ownership of accounts , which makes agents safer, more predictable, and less complex. You can think of this as turning your accounts into Google Drive folders.

An application can create sub-accounts, where you can decide who has access to which resources at what time . This allows these AI agents to easily manage the complex account portfolios of thousands of users at once , and investors can access their funds anytime. This is a small map. I am speaking in consideration of time. But, you have an AI agent that can manage many different accounts and end users can also access those accounts.

One of the exciting aspects of this is what happens to that AI agent when you decide to stop paying the credit , right? I mean, the Deep Sea has become much cheaper cheaper , so it might not be a big fear , but it's still a possible reality. The great thing about this is that even if the AI ​​agent disappears completely , you can still access the sub-account that that AI agent was managing for you through Flow's unique account management . And you can transfer all those assets to your custodial account , which is very exciting . Finally, there is something called 'capabilities' in Flow.

You can think of this like AWS IAM . This allows you to grant fine-grained permissions to resources, activities, and assets. For example, the program can specify that "this asset can be loaned for such a long time or that these users can access it under certain circumstances ." This allows you to precisely control what and where your AI agent can access . Okay, without further ado, I'll hand over to my colleague Boha to show you a little demo we have for you .

Hello everyone. I'm sharing my screen. Oh, here's my console, my guide, and VS Code. Yes. First, here you can find a repo we call 'Eliza Flow'.

In fact, it is a much lighter version of the Elisa framework. We have already provided a link to this on our hackathon event page . It is under the ' . It is under the ' Unflow' framework , and as I said, we have introduced many features for Flow in this framework have introduced many features for Flow in this framework . Furthermore, we offer an account portal, which means a real account management agreement for agents.

If you are new to Elisa, I recommend taking a look at their official Elisa Framework Readme for guidance . But this Elisa Flow version is so lightweight that you can directly use this framework for everything like a plugin . It is compatible with the original Elisa framework . So if you have already built your agent with Elisa , you can continue to use your own framework . But, if you want to build unique features on the Flow blockchain using account management , I recommend using Elisa Flow to start your project .

Here is a guide. I'll take a quick look at it. There are some basic requirements, you can clone the file and start from there. One thing you should note is that there is a dependency on the Elisa framework, so you will need to use module update to complete everything , as you can see here , sorry. You need to update.

If you already have it installed have it installed , you need to update. What is happening? Some modules are updates and this is in a recursive manner. What is recursion? What is recursive?

Recursive to update. Yes, I have already updated . Use this command to clone all of the Elisa dependencies . After that, don't forget to install everything under the PMPM store . But remember, you need to add 'no frozen lock file' because dependencies are updated frequently .

So, sometimes you want to add your own plugin , but it doesn't have a PMPM log file . But you use it directly . It may tell you that some files are missing from the log file . So, if you use this 'no frozen lock file' , it will always install successfully. And, yes, when you encounter some errors, you'll see it here .

You just need to follow the guide. This command flow is fundamental to the blockchain . You need to establish a Flow Dependent Agreement . Before that, you can watch it here. The Flow CLI here is Flow's dedicated command line tool.

If you are using Flow on a Flow framework, it is recommended to install this first . Follow it exactly . Install Flow Contract dependency . You will see the Flow contract being downloaded directly from the mainnet and saved to your local folder . This is an import folder.

Here it is. You can see that the import folder here contains all the Flow- related smart contracts . I think it will take a while to finish- yes, I already downloaded it , so everything looks fine. Then, just use 'PMT build' to create the structure, it's very easy. That's all.

In this build I have removed some files that are not Flow related . I have removed those non-Flow files , so it will save you time if you create only in Flow . But if you build on other blockchains , for example ICP, and want to use other blockchain's plugins , remove this line and it will compile and build all ICP files. But, if you are a Flow Delegate Agent , you can simply use 'PMP build' directly and it will directly create each file for you. After that you will finish the setup , set up the framework, copy the new environment file and select your model .

In fact, recommend choosing the best model to start your AI agent with recommend choosing the best model to start your AI agent with , as the model's intelligence will have a huge impact on its operations . In general, initially I used a small local model like Olama on my laptop , but that model was not smart enough to get the right answer . Therefore, I recommend using powerful models like Open AI or Deep Seek for intelligence . They will be very helpful for your AI agents . And I choose Deep Seek .

But yes, Deep Search's API seems a bit slow lately , but I'll try it later for a full test . If you choose a model , it will be very simple . Here is a file we created for a little Eliza. This is a JSON file created by Ali. This is a descriptive file containing her character information.

But remember that when starting the ELISA build , you need to check the schema of that feature file very carefully . Therefore, it is best to copy the ELISA characterization file as is and modify everything yourself , as the fields are very strict. Here you can customize a lot of message templates and their personality . But here's the most important part. This is your your model.

I have selected Deep Seek , and here is the secret information about your Flow wallet . You need to set up your Flo wallet . Remember, this is for the Agent Wallet only. This is not a testnet. You will need to set its private key and Flow wallet address.

But remember that it is only for the test agent wallet . This is not for the user wallet. I'll talk about account management later , but here are the plugins. A plugin is something you need to load. You just need to post the name of this plugin.

Also, this plugin is Flow- based based . The plugin package I wrote is ' package I wrote is ' common' and its name is ' fix AI common'. I'll paste the plugin name here. All of these are a very important part of your character. Add your plugin, add your Flow wallet, select a model , and you're done.

That's all for the agent setup. Then, if everyone has it, you can run the code to start your AI agent. If you are in development, I recommend using the 'npm start debug' command . This will show the additional information needed by the agent. I have some orders , here is one.

Start debugging using the properties file JSON. Loading and finished. It has been started been started . As you can see in the post here, I have chosen the Deep Seek model for the character model , and I have recorded several actions here . It is a specific action derived from a Flow-based plugin.

This is a common function coming from the Elisa remote . I will tell you the details of this action later. But if you look at the last sentence here , it says that if you want to communicate with your agent, you can use PNP start client to start a client to access the agent. So I use this window to start PNP . After that you can create a window.

You can create, here it is. A very simple client to operate. Here is a chat box. You can chat with your AI agents. Let me ask you a very simple question first.

And tell me my wallet information . That's not the agent's wallet information, it's my wallet information. So, you can see the lock here. Since I'm using debug , I can see that the prompts for agents are very detailed and have some examples. Here are all the prompts .

Here it will say, " it will say, " Okay, your message is 'flow' information ." We will run the 'get user accounting' action . Then you will see here. Here is all your account information . This is an action I created in advance in the 'Alteryx flow' plugin, a generic plugin.

What we see here works for this . You can see that this account is configured as an agent account in the JSON character file, we can check the 'flow' again . Then check that this is our agent account address . Wait a moment . The website is a little...

hello. Oh, sorry. Test this. Hello. This is our agent address, but you will see the answer is your ' flow' wallet address.

Here is another 'flow' wallet . This means that the agent's wallet controls your user ID , the agent controls your user ID and assigns you another 'flow' wallet, which is a dedicated wallet for your user ID . This is a different wallet. This is our on- our on- chain account management and everyone uses it . Using this framework and protocol , each user can chat with your agent and be assigned a different wallet.

One wallet for every user. All operations you perform with the agent are based on your wallet instead of the agent wallet . This is what we call on- we call on- chain account management . And here are some examples , I will explain the detailed steps later . This is a common plugin, we remember...

sorry. Didn't see it . Initially, we start our agent with RC5 'flow' exclusively . Five special actions. These are those five activities.

These five functions are pre- created created . Ensuring that the user account exists , i.e. creating a user agent for this user ID . Here you can see that we are creating based on the user ID for each user, using it as an identifier.

Create a wallet based on this user ID and get your wallet information. Also, here is a simple way to get the on- chain flow price and get the flow price and token information using an oracle query . You can get all the Flow token information, just send the smart contract address and it will provide all the details like symbol, decimal, and price in Flow. This is a different exchange from other blockchains. This transfer is to your account instead of the agent account .

Here you will see that we are using the user ID. Remember the user ID as the controller of the wallet, transactions where you change the token will take place through the account service. You need to provide your user ID as the key parameter, and all changes will come from the user's wallet instead of the agent wallet. This is our native account management service. Another feature we are introducing is our bundled exchange.

In fact, the exchange is very simple at present. This can be used not only for Flow, but also for Cadence and EVM side . For the EVM side, we use Cadence to wrap the EVM transaction , so that we can do multiple transactions like we do multicalls in Solidity, and you can add additional conditions and logic to your scripts to do complex transactions . For example , here's an s-it's not just a simple flow token information query, you can also do transactions beyond token information, decimal, contract, etc. If you want to combine all the pricing queries into a single API call, along with basic information, you can use a cadence file .

Sorry, this is ERC-20 token information, here is some. Price, I'm skipping many specific smart contract addresses , think of this Cadence file as set of query information like JavaScript , you can add it all. All API calls calls , in this single file. And here is how to use this to get information use this to get information , you can get EVM token information. And this execute transaction can be used .

This command only uses the Execute transaction to get EVM information uses the Execute transaction to get EVM information . And you can write all the scripts in this CDC file. And you can add everything, the information you want to query or any transaction you want to execute, into a single file . We call this a batch transaction or batch script query . You can get not only the token pair, but also the token price, get all the decimal information, organize it into a structure , and then return it to a file .

So, if you are an EVM developer , you can also use specific features of Cadence and Flow to enhance your development experiences . I think these two key features are the most exciting part of our AI agent development , yes. So, this is about to end Here's everything I wanted to share, I'd love to share more information about this . I 'm keeping an eye on the time, we still have two minutes . I shared a link in the chat .

Join the World Tour channel , namely Flow World Tour. If you type it into Telegram , it should work . That's where we provide feedback. People generate different ideas . If they have any questions about our rewards , or if they would like to receive pitch training.

To be honest , take advantage of this because our team has about 15 first place winners in hackathons individually . People who have started unicorn startups or were founding team members of unicorn startups or started their own startups . There are also trainers who have been trained by TEDx trainers and have spoken in front of thousands of people . So, this is a great opportunity to create a winning pitch at the hackathon and find the right market fit for your idea. Find out what pain and problem you are actually solving ?

Who are you solving that problem for ? What will the future look like What impact will result from what you create ? These are the main things you need to think about when it comes to our rewards or when you are hacking in general. What are you solving solving ? What do others work on ?

What are the shortcomings of the alternatives people have ? And how do you make a big impact ? Try to provide the necessary statistics for that . Try to show why what you create is important . Plus, oh, I saw a cable.

Yes, that's right. Then, there are also partner channels on the ETH Global Discord . Join it. We are operating there too . Thank you very much to everyone who participated.

I think it was very useful. Thank you very much, Bohau. Again , we are here if you have any questions . Good luck with your hacking . Okay, let 's start the show now.

Automatic transcript — names and jargon may be misspelled.