OKX × Bloom: I Gave an AI Agent My Wallet

We ran a Bloom seminar and workshop with OKX Wallet, a global Web3 super app, and a speaker who used to work at OpenAI.

Honestly, since this was the blockchain world, I hesitated a bit when we first started talking about working together. Above all, I wasn't sure it fit the feel of our community. Then the special speaker, a developer who used to work at OpenAI's headquarters, came over from the US to Korea for a short stay, and when we talked, he told me AI agents and blockchain are much more connected than I'd think. He had also worked on related things in the past, so he said he could explain the link between AI and blockchain well.

Even so, I don't know much about deep tech like this either, so I specifically asked an AI developer at our company, Cliwant, to moderate. I figured there would be a lot of development detail, so I asked them to explain it in a way that would be easy and fun even for someone like me who knows nothing about it. In the end, a topic I assumed would be too hard and unfamiliar turned out to be a really interesting night with a lot to learn.

Agents that act and pay, past search

We set the theme of the event as AI agents. AI has moved past searching and doing research for you. It's entering an era where it actually takes action and even makes payments. Given that, we wanted to talk about how we're using agents, and how far agents that really act, track, and pay can go. I could tell from the sign-up stage how much interest there is in where AI and Web3 meet. Far more people applied than we expected, so much that I was sorry we couldn't fit everyone.

This time we also added time to build, not just listen. We asked people to bring laptops, and since the hands-on part needed tokens, we decided to airdrop tokens into every participant's wallet so they could actually make trades. If you make money, you keep it. If you lose it, it's not a big amount, and it's a good experience either way. Since it could be hard to do alone, we set up groups of four so that people who knew what they were doing could support the others.

How finance companies bring in Claude Code

The opening talk was from Yongho Choi, a tech evangelist at AWS, which sponsored the venue. This was his second talk at Bloom. His topic was how the financial sector uses coding agents, and the case study was Bithumb, a major Korean crypto exchange.

The background alone was interesting. AWS has strong partnerships with Anthropic and OpenAI, so the Claude models behind Claude Code run on AWS infrastructure. Amazon Bedrock, AWS's generative AI service, lets you pick from many models instead of pushing one proprietary model, so both Claude and GPT are on it. The key difference is who processes the data. On Bedrock, your data is processed inside AWS. If you use Anthropic's API directly, it's processed in Anthropic's environment. And if your servers are already on AWS, a feature called PrivateLink lets you use the Claude model in the same data center without going over the internet at all. For heavily regulated places like finance, that's a deciding factor.

Bithumb's situation is one a lot of companies go through. Developers wanted to use Claude Code or Codex to be more productive, but the company couldn't provide them because of security regulations, so people started using them on their own with no oversight. When the company surveyed internal demand, there wasn't a single developer who wasn't using them individually. From the company's side, that's a shadow IT risk: you don't know where your code is going over the internet. So Bithumb set five security governance standards (data privacy, isolation, network, access control, and auditing and monitoring), and after research it brought in Claude Code on Bedrock. It's tied into internal authentication, so you have to verify your identity with MFA to use the tool. A central LLM gateway handles every request, which allows token tracking, audit logging, and filtering. Before anything reaches the model, guardrails filter out harmful questions, confidential information, and personal data, and logs are collected in one place, where they even run correlation analysis on security vulnerabilities. They also turned the process of getting trained, applying, and getting approved into a platform. And so that the skills individuals build don't stay as one person's knowledge, they built a skill hub where anyone can reference the well-made ones.

As a result, shadow IT went away, developer productivity went up a lot, and they can now analyze in detail how many tokens each team used. They can use the best tools in a controlled environment and switch between them anytime. The questions were sharp too. Someone asked whether it's technically guaranteed that AWS won't use model parameters for its own purposes. He answered that the contract structure keeps even AWS employees from accessing them, and that this is certified by an outside third party, not just verified internally.

Ten emails to Sam Altman

The first fireside chat was moderated by Yeryeon Seo, an AI engineer at Cliwant, the company I run that hosts Bloom. The guest was Younghoon Kim. He introduced himself as someone who worked as a developer at OpenAI for two years, left this past April, and is now trying to start a company. He joked that when he heard OpenAI was going to IPO, he rushed out thinking he'd better try it himself. He's back in Korea for about a month to reapply for his O-1 visa, and said he'd wanted to come to Bloom ever since he was in the US. When he heard this event would cover blockchain, agents, and security all at once, he thought he could tell his story in a fun way and said yes right away.

His path started with a master's in financial engineering at ETH Zurich in Switzerland after college, where he researched how to apply existing banking regulations to DeFi. During his master's he founded a startup that detected AI-generated images and text. Stable Diffusion and Midjourney had just come out, so there was demand from places that handle artwork, and from the College Board, which runs AP exams in the US, to tell whether a student's final project was really their own work. But what he realized doing this was that the core question isn't real or fake. It's who made it. Which artist made it, and through what process: what's called data provenance. That view lined up exactly with OpenAI's Trust & Safety. The work he ended up doing at OpenAI was figuring out what intent people using the API or ChatGPT had, and whether there was harmful intent or a risk of financial fraud.

The story of how he went from intern to full-time employee was the best part. He built an email system for notices that regulations require be sent to accounts sanctioned for policy violations. He tried sending a test to his own organization, but the email kept not arriving. Assuming it was the kind of temporary glitch every developer knows, he kept clicking until it worked. Around the tenth email, it turned out he'd been sending them not to his own internal org ID but to a legacy org ID used by leadership like Sam Altman and Greg. He only found out when a director rushed into a public channel and asked, "Is that you sending these right now?" He laughed that it definitely got him noticed, and that since bad publicity is still publicity, having people remember his name probably helped his conversion.

Individuals use it well, so why do teams struggle

The main topic was the problem that led him to start a company. OpenAI is the organization that should use Codex best of all, and internally people really do use it without limits and have deep know-how. How to use Codex or Claude Code as a personal assistant is already well established. There are skill files and harnesses everywhere. Ironically, because each individual uses them so well, problems started showing up in collaboration. People edit the same files and the same project with their own workflows, so agents overwrite each other. One team prefers verbose documentation with two or three lines of comments for every line of code. Another team wants to throw out everything unnecessary and let the code speak for itself. When those two teams work on the same product, it's write, delete, write, delete, over and over. He said he saw this far too often and started the company to fix it.

There were also problems specific to frontier labs. To ship a single AI product, you have to clear a huge amount of compliance. For example, to release Codex you have to put up guardrails so users can't use it to make biological weapons. If someone buys avocados and cabbage at a well-known American grocery store and asks how to culture anthrax, and they actually do it by following the answer, OpenAI is responsible for helping a terrorist. Including absurd scenarios like that, 20 or 30 teams have to come in and coordinate, and it goes through many hands, from legal to product operations to user operations. So even after the code is done, he said it takes three to six months to launch.

The unit of work is the workflow, not the document

This is where the picture for Planar, the company he founded, comes from. The reason we had to use tools like Google Workspace, Notion, and Jira is that the unit of work between people was the document. We wrote PRDs, made design docs, and collaborated by sharing them. He called that the old way. In the age of agents, he said, the unit of work should be the workflow, meaning the loop, instead of the document. In practice it may still be expressed as a document in the form of code, but he wants to own the problem space of what form defines that loop best.

His first tool for this is the forward deployed agent. A lot of companies these days have Palantir-style forward deployed engineers, but not every company can hire people like that. So instead of a person going in, holding meetings, and building tools, you integrate an agent into the workspace and have it discover everything on its own and even build and hand back the workflows. It connects through MCP and APIs, scans continuously in the background, and hooks into Claude Code to find which workflows the agents running at that company use and what work keeps repeating. For now they're targeting industries with lots of repetitive work and a high return on investment, like finance, legal, and accounting, and he said the ideal customer is a company that's already comfortable with agents. If you have to start by convincing a company what an agent even is, there are too many hills to climb. When Yeryeon pointed out how hard it is in Korean AX (AI transformation) work, where practitioners first have to explain why things don't just work with one click in Claude, he agreed that Korea seems like a tough place to find that ideal customer.

On the big question of what people will do once agents handle most of the work, he named ownership and delegation. People talk about zero-person companies, but in reality there are points where a person has to be involved, and once you're juggling five Claude Code sessions, you feel deeply that you only have one brain. In an organization running many agent loops, when someone has to ask which agent produced a piece of work, the one who has to answer is the person who built that agent. So the most important thing is to clearly tell each other where ownership begins and ends: my agent has this goal and runs in this codebase. That's also how you avoid the overwriting problem, he said.

The part about agent identity was interesting too. If agents are going to really automate our lives, payments can't be left out, so how do you give an agent an identity? Right now the common approach is to hand over the owner's own identity as is. He pointed to hierarchical deterministic key derivation, a cryptographic technique long used in blockchain. Delegating authority to an agent means the same thing as deriving a key from your wallet, so with this technique, when an agent spins up a subagent, delegating authority from the main wallet to a sub-wallet can be handled programmatically. He said this approach will really help once agents get payment authority.

On a two-pizza team, just fight it out

The Q&A was packed too. Asked why the workflow becomes the new unit of work, he said it might sound like wordplay, but if you have an agent running 24 hours a day, there will always be repeating workflows, and he's interested in how people define, edit, and share those together. Asked how to resolve clashes of taste in multi-agent collaboration, he was characteristically blunt. If you're small enough to be fed by two pizzas, drop the titles, argue it out, and go with whoever wins. He ran into exactly this problem himself during a weekend hackathon with his roommate, who is also his cofounder. Now, before building the product, they put their skill files and agent config documents in front of the two agents, have the agents talk it out first, merge everything into one, and then start. They split up the work by deployment unit: mobile, backend, and the connection between them, so each piece can be deployed on its own.

Asked how to deal with strict financial governance in the US, he said honestly that he doesn't think there's an answer. Even OpenAI uses these tools with extremely limited time and scope, for example five people on a specific team for one month, and he said it's closer to using governance as a moat to eliminate competitors than to solving the governance problem. One attendee said they were doing AX at a legacy company with a local LLM. He said that work is a lot like what forward deployed engineers do, and that his bet is that not every company will be able to hire people like that. Finally, asked how he uses tools these days, he said he subscribes to every vendor in rotation, but there's a common pattern: separating an inner loop from an outer loop. The inner loop is where you build the features you want one by one. The outer loop catches the quality loss that comes with it, so-called AI slop, like code conventions breaking down or design logic drifting. He keeps simple skills ready to throw in whenever he spots a problem, and one loop running 24 hours a day in the background is always catching those.

In the agent economy, trust becomes currency

The second session was a panel run by Taylor Choi, who leads growth at OKX. In a post ahead of the event, he wrote that watching the AI industry grow keeps reminding him of how Web3 grew. A new technology appears, builders rush in, countless projects pour out, and in the end the ones that survive are the services people actually use. A few years ago at an event, he said AI would lower the barrier to entry for Web3. Now that cases of AI agents handling complicated on-chain processes are starting to appear one by one, the most interesting thing to him is that what he said back then is becoming real. He's also the one who set up this collaboration, saying he'd wanted to work with us someday ever since he saw our content back in the Claude Blue days. The panel included Qiao Zhao, a senior product manager at OKX AI, and Ye, head of product at X-Agent, which lets people build agents in natural language. The panel was in English with live captions, and it opened with a laugh when Ye greeted the room in Korean with "annyeonghaseyo."

Asked how they define the agent economy, Qiao Zhao said the heart of an economy is providing services. Just as value is created and moves when one person provides a service another person needs, the agent economy is one agent providing a service another agent needs. What OKX wants to build isn't the smartest AI in the world. It's the infrastructure of that economy: a platform where every agent can find work, get hired, and get paid. Ye defined an economy as a system that coordinates resources toward goals and named three layers their team treats as equally important: security, meeting the goal, and quality. When a user builds an agent in natural language, is it safe, does it meet their real need, and does it get a better result than other tools.

Asked why we need agents at all, Ye had a fun analogy. Agents aren't air. Without them, the sun still comes up tomorrow. But you can reach your destination without a car, and walking five hours at 95 degrees in midsummer still isn't the smart way to do it. Agents are the tool for reaching a goal in the best and easiest way. Ye joked about being greedy and wanting an agent that brings in a million dollars every day. Qiao Zhao used apps as the analogy. To talk to a friend you open a social app, and to order something you open another app, but with an agent you say what you want and it handles everything. Once you've had the feeling of setting a goal, going to sleep, and waking up to find everything done, there's no going back.

Asked what becomes valuable when anyone can build an agent, both named trust. Just as you look at someone's track record when you hire a person or buy something, agents need a record of results and a reputation before other agents will hire them. On how much control to hand over, Ye said this question has a right answer: least privilege. If an agent manages your Gmail, give it only the minimum access that job requires. Ye brought up an incident that made the news earlier this year. An executive at Meta let an agent manage their email, watched their messages get deleted right in front of them, and could only stop it by shutting off the laptop. If accidents happen even to someone that careful, Ye said, every AI product should put first the fact that a single big accident can take everything from a user. Qiao Zhao described two lines they draw. One is code, meaning setting what the agent should do. The other is money. But if you never give the agent a wallet, a person always has to do the last step and the loop never closes, so you need a balance: set a limit on how much the agent can spend and let it finish the job within that.

Their predictions differed. Qiao Zhao pictured a world where, three to five years from now, there are no app stores left, only agent stores, and you just state your requirements and the agent you need shows up. Qiao Zhao also predicted that agent payments will come much sooner than people expect. Right now agents can answer and code but can't pay, so a person always does the last step, and that piece will soon be filled in. Ye said everything is moving too fast to predict, but was sure of one thing. People don't want to build agents. They want the results agents bring. So in the future, the act of creating an agent will itself disappear, and we'll live in a world where you throw out a problem and it's already solved.

Building an agent with a wallet

The last 100 minutes were a workshop. The OKX AI team ran it, and Taylor interpreted. The steps were clear: install OnchainOS, make a wallet for the agent, put money in that wallet, and write and run a trading strategy.

OnchainOS is a skill you install on your agent. You copy one command from the website and paste it into Claude Code or Codex, and once it's installed, the agent gets a wallet, OKX market data, trading functions, and the ability to make payments. The wallet is fully self-custodial and you create it by logging in with email, so nobody, OKX included, can touch the money. It's built on AWS's trusted execution environment, so you can recover it with just your email. When participants submitted their wallet addresses through a Google Form, we airdropped 10 USDT, worth about 15,000 won (roughly $11). Sixty-eight people submitted. Of course, hands went up all over the room with CLI errors, and the room burst out laughing at Taylor's tip: if Claude won't listen, just firmly order it to do it.

The highlight of the hands-on part was the trading agent. Decentralized exchanges these days trade tokenized stocks as well as crypto, and popular stocks like Nvidia are already listed. The demo used tokenized SpaceX shares. You ask the agent, as if you're chatting, to monitor real-time SpaceX news, judge whether it's good or bad news, buy and sell accordingly, and even build a profit and loss dashboard, and the agent works out the strategy and builds the bot on its own. It runs on a chain with no fees, so you trade with 100 percent of your money, and you can choose whether to approve every trade before it executes or hand everything over. The presenter made an important point. If you don't tell the agent your risk tolerance, like your budget, how much to put into each trade, and where your stop-loss and take-profit points are, you won't get a safe strategy. Unlike people, who get swept up by emotion and panic sell, an agent leaves emotion out and trades only on the news and the strategy. The line that you can leave the agent running in the background and go live your life sounded like the theme of the whole event.

The last thing introduced was the OKX AI Marketplace, which had launched just the week before. You list an agent you built, and other people, or even other agents, find that service, pay, and hire it. In Taylor's words, it's the Kmong of AI (Kmong is a popular Korean freelance marketplace, like Fiverr). It's not just trading. Anything goes up there, from an agent that analyzes World Cup win odds to a health agent that plans your meals. If you tell your agent, "Find me an agent on OKX AI that can build a SpaceX trading strategy," it pays less than a dollar and brings back a report from a specialist agent. An economy where agents hire agents is actually running. OKX founder Star Xu has even publicly promised to give at least one bitcoin of his own money to the first solo entrepreneur who reaches $1 million in annual revenue on this platform. That person could come out of this very room.

Late into the night, groups of four built their own trading agents and laughed as they asked their agents about their wallet balances. In an age when agents work on their own and even spend money, I wanted to know what people should still hold on to. Putting together the answers I heard that day, it came down to goals, limits, and the boundaries of ownership. Deciding what the agent should do, deciding how much it can spend, and being able to say this agent is mine and I'm responsible for it. You can give the agent the wallet, but those three things are still the human's job.

Originally written in Korean on Brunch, July 10, 2026.