I hopped into a taxi in Bangkok last week and the driver, a man north of fifty, spent the ride telling me what he was building with AI. Not complaining about the economy. Not asking where I was from. Telling me about his project. I’ve been turning that over ever since, because it isn’t an isolated thing. For weeks now I’ve been scanning event listings in whatever city I land in, and the pattern is hard to miss. It isn’t pitch nights anymore. It isn’t another fireside with a fund manager. It’s vibe coding meetups, agentic AI sessions, AI trainings. Paid attendance, no walk-ins, speakers who’ve shipped real apps. KL has them. Singapore has them. Bangkok and Manila have them. Go on Lu.ma or Eventbrite right now and there’s probably one happening in your city this week, maybe two. I know this firsthand because I run some of them. I host AI salon events in Bangkok, and I’ve watched the rooms change. So while the rest of the startup world argues about whether funding is back, looking at numbers that are frankly pretty dismal, there’s this whole other thing happening in cafes and malls across Southeast Asia. Regular people are learning to build software by talking to a machine. Why I trust this one I dismiss most AI hype on reflex. My feed is littered with slop, articles that read like they were generated by the thing they’re describing, people calling everything the future. I scroll past it. This is different, and the reason is simple. People are paying to show up. And it’s a different crowd than I’m used to seeing at startup events. University students and fresh grads who can see the job market tightening and are choosing to get ahead of the curve instead of waiting it out. Founders who can’t afford a dev team. Marketers. People with an idea and no technical co-founder, who a year ago would have been stuck with that idea trapped in their head, never seeing daylight. This is the no-code, low-code movement, upgraded and supercharged into the current AI era. The category has a name now: “vibe coding”. I’m not a fan of the term, all that talk of vibes and feel grates on me, but it’s the vernacular, so I’ll use it. You describe what you want in plain language and the AI writes the code. That’s the whole thing. I do it myself. I’ve used AI coding to replace most of our software stack. Thinking back to the friction of a couple of years ago versus how good this is now, and then projecting forward to how good it’ll be as the models keep improving, is genuinely one of the more interesting arcs I’ve lived through as an operator. From apps to agents, which is where it gets serious Building an app is one thing. The next rung up the ladder is building an agent, and agents are a different animal. Most people, once you get out of the tech bubble, still picture a chatbot. You type, it types back. You ask, it answers. A better Google. That’s generation. It makes text, images, words. An agent acts. It doesn’t tell you how to clear your inbox, it clears your inbox. It books the meeting. It sends the email. It runs commands on your machine. It talks to other software and gets things done with barely any input from you. That’s the entire ballgame for risk. A chatbot needs a human to type every prompt. Every harm one causes still started with a person asking for it. An agent can plan, decide, and act on its own initiative. It can cause harm nobody asked for. I want to be clear that I’m bullish on this. Hugely. But being bullish and being measured aren’t opposites, and the risk side of this deserves honest airtime. Two examples everyone in the open-source world is talking about. The first is the lobster: OpenClaw. It went viral the moment it dropped. It connects an AI model to your messaging apps and acts on your behalf, books things, browses, runs commands, manages your house. People pulled their old Mac minis out of drawers to run it. Apple caught the wave and nudged the price up. It is not a Southeast Asian product, and we should be honest about that. It went viral hardest in China, which has been well ahead on the open-source movement. Southeast Asia needs to kick into gear as a fast follower, even when we’re not the origin. The second is Hermes, out of a US research lab a few months back. What makes it different is memory. It lives on your own server, runs all the time, and gets better the longer you use it. It remembers what you told it last Tuesday. It writes down how it solved a problem so it never starts from scratch again. By this month it was the most-used agent out there by some measures, hundreds of billions of requests a day, hundreds of thousands of developers piling in within three months. Here’s the part that should make you pause. Three separate security audits this year found malicious code hiding in the add-on skills people share for these agents. Think about what that means. An autonomous thing, running constantly, on your own machine, with access to your messages and files and maybe your ability to spend money, pulling new abilities from a community marketplace that’s already been found to contain things designed to hurt you. That isn’t a future problem. It’s a this-year problem, and it’s happening on hardware people own, in their homes, outside any IT department or compliance check. A friend who’s far sharper than me on this put it well. Permissioning an agent is like onboarding a new intern. You give them enough access to act, but not enough to break things. If humans are entities of action, we have to treat agents as entities of action too, with the same scoping and the same limits. The catch is that getting that right still takes real technical skill, and most of the people downloading the lobster don’t have it. So who’s writing the rules Surely someone’s regulating this. Here’s where it actually stands, and the answer is more interesting than “nobody is.” Three big global players, three different postures. The US is actively deregulating to keep its lead, tearing up the old safety rules and trying to stop its own states from making their own. The posture is get out of the way, though there was an executive order floated recently that would have made new models notify the government before public release, something closer to how the FDA approves a drug. It got paused, not signed. We’ll see. Europe, true to reputation, has the most serious regime, and just this month agreed to delay the hardest parts, the high-risk rules, by over a year. Competitiveness pressure. So even the strictest regulator in the world is loosening its grip right as agents arrive. And China is the strictest in practice and the only one already acting on agents specifically, real enforcement, thousands of non-compliant services shut down. Telling, the country where everyone installed the lobster also told its own government agencies and state banks not to put it on work devices. The adoption champion got nervous about its own craze. Even the deregulating US quietly started building standards for autonomous agents. So nobody actually thinks this is fine. Everyone sees the gap. They’re just moving at wildly different speeds. Southeast Asia is that same story compressed into one region, running at three speeds. Vietnam, maybe not who you’d guess, has the only real binding AI law here, passed late last year, enforced since Q1, risk-based with actual prohibited uses. It tracks, given how much of the region’s developer talent sits there. Singapore did something very Singapore: the world’s first governance framework built specifically for agentic AI, detailed and thoughtful, and deliberately voluntary. No teeth. The bet is give industry sophisticated guidance, remind everyone they’re still liable when their agent screws up, and keep the innovation onshore. They’ve already refreshed it with case studies from the likes of OCBC, Tencent and Workday. A living document, which is the right call given the pace. And then Malaysia, where I’m based, sitting on one of the most aggressive agent rollouts in the region, with its actual rules still in draft. Not here yet. Here’s the whole thing in one line. Everyone, globally and right here at home, is regulating the last war. The last war was chatbots generating bad content, the stuff you can ban after it spreads. We saw it when Indonesia, Malaysia and the Philippines banned Grok over deepfakes, including images of children. Three countries, fast, coordinated, and fully deserved. But that’s the model: react after the harm, fold quickly. And every one of those images still needed a human to type the prompt. The next war is agents taking bad actions on their own, because the black box decided that was the thing to do. That war is already shipping. Through anonymous downloads, onto personal machines, learned at meetups across the region, in a place where exactly one country has even a voluntary framework and the country with the biggest rollout is still drafting. We’re banning the thing that needs a human to ask. We haven’t started on the thing that doesn’t. What I keep coming back to I’ll be honest, I don’t have a clean answer. Part of why I raised this is that it was a quiet news week. But the bigger part is that I can’t stop noticing the trend, and I doubt I’m alone. If you’re a CISO or a CTO or sitting in a compliance function, you’re already living this, because the whole enterprise is integrating more automation and more agents by the month, and the risk side is going to drag a regulatory environment into the room whether we invite it or not. It always does, the moment a technology touches enough of society. So it’s worth thinking now about what that reaction is likely to look like, instead of being surprised by it. But I keep coming back to those meetups. To the rooms full of people building. Because that’s the real story, and it isn’t happening in a lab or a boardroom. It’s happe
