OpenClaw is the New HTML

I know by the time this is in front of you, OpenClaw has already gone from the thing everyone was talking about to the thing everyone is moving past. I've been observing and monitoring until the dust settled, and here are my thoughts on OpenClaw.

For those who missed the OpenClaw hype: OpenClaw is the fastest-adopted Agent Framework in history. I know people now call it a "harness" — after context engineering, after prompt engineering — but I'd stay old-fashioned. The term "framework" is good enough for me to cover what we need.

So OpenClaw lets you run a Claude Code-like agent on your own Mac mini, accessible via IM, not just from a terminal. Or, for many people who had never used Claude Code before, OpenClaw is a much more powerful tool than ChatGPT or Claude — because it can actually create things or operate on your computer via Telegram, just by talking.

The whole arc was four months: launched in January, post-hype by April.

For many people, it burned bright and burned out. And within a week of installing it, a quiet question started showing up on the subreddit — and stayed: "I set up OpenClaw but I don't know what to do with it." It was named, papered over by tutorial lists, and never quite resolved.

But this is not a problem with OpenClaw: when you have HTML, what website do you build? If you have no idea, that's not a problem with HTML.

Yes, HTML. I think now is the proper time to rethink: what does this whole seance mean? Where is it heading? Is there a future, or just a dead end already?

My answer is that the framework might evolve, but OpenClaw started something categorical and changed something forever: OpenClaw is the new HTML. When I say OpenClaw is to the Agent Economy what HTML was to the Web Economy, I don't mean this framework will last for 30 years, nor that it's the best agent framework of the moment.

Just like HTML wasn't the best document format either. By any rigorous standard, HTML was worse than its alternatives. SGML was more expressive. TeX was more precise. Xanadu's hypertext model was more theoretically complete. HTML was a sloppy subset of SGML with a handful of tags, no formal schema, and notoriously forgiving error handling.

Yet HTML started the whole Web Economy. Yahoo, Google, Facebook, Amazon (and AWS) — the size of the Web Economy reaches roughly $16 trillion a year, about one-sixth of global GDP. And what Agents will address is not just the Web Economy — it's the labor market itself. Roughly $30 trillion a year, globally.

And HTML didn't get there by being the best — it got there by not being the best. OpenClaw is very similar in this regard, and here's why.

1. Permissionless

Anyone could create.

For HTML: anyone who could type <h1>Hello World</h1> into a text editor — and if they typed it wrong, the browser would render it anyway. Before that, you needed to be a sysadmin who understood FTP, file permissions, server configuration, directory structures. The population of people who could put something on the Web went from thousands to billions — overnight, in principle.

For OpenClaw: anyone with a problem worth automating. Before that, you needed to be an ML-fluent developer writing LLM API calls, designing prompts that survived edge cases, wiring up tool use, managing memory across turns, evaluating outputs. The barrier dropped by an order of magnitude — and so did the question of who counts as an agent builder.

Now you talk to OpenClaw like you talk to a human, and it reshapes its own behavior. Anyone can build.

Publishing required no permission.

For HTML: no editorial committee, no curated store, no central directory. You put your HTML somewhere accessible, and anyone with the URL could find it.

For OpenClaw: no App Store, no GPT store, no vendor approval. You build your agent, and anyone can run it on their own machine.

Publishing wasn't granted. It was done.

Nobody owned it.

HTML the standard belonged to no one. Browsers rendered whatever they received. There was no vendor to ask, no platform to apply to, no API to call.

OpenClaw the framework belongs to no one. The runtime executes whatever it receives. There is no vendor to ask, no platform to apply to, no API permission to obtain.

Anyone with the energy to build, built. Anyone with the energy to build, builds.

When "who can participate" changes by orders of magnitude, you don't get more of the same thing. None of the forms we now associate with the Web were designed. E-commerce, social media, user-generated content, the platform economy, the creator economy — they emerged. Because participation was permissionless.

The Agent Economy will be made the same way. For the same reason. So Agent is not an app — constrained by build frameworks and gated by distribution channels for publishing.

Agent is not the new app. Agent is the new Web.

2. Build & use, at your own discretion.

Why don't we have a similarly "powerful" personal agent from Google or Apple? Because an agent with enough Agency to act has real-world consequences. Not just chat. Guess who gets held responsible when shit hits the fan?

OpenClaw is as flawed in security as HTML was at its birth. The safety stakes are even categorically higher than anything HTML ever had to deal with, because OpenClaw agents act at the user's command. They handle private data, make decisions, send money, delete files, take steps across services.

OpenClaw can come into daily use because being open source means no one owns it. It's not a service. No profit, no responsibility. You must use it at your own discretion — take your own responsibility.

But this is where Open Source stops being a virtue and becomes a requirement.

You can audit what an agent does. You can restrict where it touches. You can modify behavior you don't trust. You can run it in an environment you control.

That's the obvious half of why Open Source matters here. The half that gets missed: this is also how a field actually matures. The Web went through exactly this arc — XSS, SQL injection, session hijacking, the whole 1990s litany of catastrophes were all uncovered in public, patched in public, learned from in public.

That's how the field hardened. The closed-service alternative is the opposite: problems stay hidden until they detonate, because nobody outside the vendor can see them. AI didn't invent this trade-off. The Web already settled it. Open won — not because open was safer in any given moment, but because open was the only path that let safety actually improve.

Then there's the vendor side.

No responsible vendor will ship the real thing. The agent that handles your accounts, your inbox, your finances, your relationships — it can't be shipped by anyone with investors, a legal department, or a board. The legal department alone would kill it. Call it the Responsibility Paradox.

3. Whose Agent is it?

There's also something deeper. Agents have meta goals — something documents never had. (Well, some people don't have them either, but that's a story for another time.) A web page has no purpose of its own; it is what it is. An agent has a purpose, and whoever defines that purpose has the agent's loyalty.

This is the Principal-Agent Problem in its purest form.

Which is another reason why no real Personal Agent has shipped yet. Because it's simply not personal.

Even when a vendor ships you something they call a "personal agent," the meta goals are theirs, not yours. It's a corporate agent rented to you. In proprietary agents, the vendor defines those meta goals — not just through terms of service, content policies, brand-safety constraints, and monetization layers, but most importantly by embedding them in the agent framework itself, where you aren't allowed to touch them — the runtime is theirs, not yours. The agent's loyalty is thus structurally allocated to the vendor. You pay for it. You use it. It is not serving you.

OpenClaw doesn't yet have a module called "meta goal," technically. The LLMs in the cloud that drive it are the layer the user cannot touch. But you can switch from one LLM to another, since this agent framework has its runtime on your machine. And in theory, you can own the meta goal in future iterations.

This is the closest thing we've had, at the consumer level, to an agent that holds loyalty to its user — whether the intent is good or bad.

Agent is the new Web.

HTML birthed the Web Economy, but the forms of the Web Economy weren't designed by HTML. They emerged. E-commerce, social media, the creator economy — none of it was on Tim Berners-Lee's whiteboard.

The Agent Economy will emerge the same way. If the answer is "anyone who can install a runtime can participate" — the network becomes permissionless. And permissionless networks produce forms that nobody designs and nobody anticipates.

Like HTML 1 to HTML 5, OpenClaw will be a Ship of Theseus — the version we're using five years from now will share almost nothing with today's beyond the name. When the category is created, new frameworks keep arriving. NVIDIA is absorbing it into NemoClaw. Hermes Agent calls itself OpenClaw's successor in its own documentation — and as of this month, has overtaken it on OpenRouter.

Now we can no longer pretend it didn't happen. The default assumption — that personal agents must come from vendors, on the vendor's terms — is gone. People's mental model has been permanently changed.

People are starting to wake up and ask: in the end, who is this agent serving? The answer is being written right now — one runtime at a time.

The open path leads somewhere — and where it leads is what the rest of this decade is going to figure out.

That's a Renaissance.