The Emperor's New Agent

OpenClaw just surpassed Linux on GitHub’s all-time star leaderboard. Let that sink in. A project that didn’t exist two years ago now has more stars than the operating system running half the internet.

The hype is real. The capability is not.

I know because I tried to use it. And what I found was a tool that is very good at looking like it’s building the future while quietly failing to build much of anything at all.

The TenantKit Experiment

My prompt was simple: give me the fastest path to $10,000 in the next eight weeks.

OpenClaw came back with a SaaS product idea: a multi-tenant kit I started calling TenantKit. It sounded plausible. It sounded like a plan. I let it run overnight.

By morning, it had barely started. Too many issues. The kind of cascading, compounding errors that don’t resolve themselves - they just multiply. I was left with a pile of half-written code and a business idea that was, on reflection, the opposite of a fast path to revenue. Building a new SaaS product from scratch is a months-long project requiring infrastructure, distribution, and iteration. OpenClaw didn’t know that. It parsed my words and missed my point entirely.

This isn’t just my bad luck. A trajectory-based safety audit of OpenClaw tested the system across 34 scenarios spanning six risk dimensions. The overall pass rate was 58.9% - barely better than a coin flip. On test cases specifically related to “intent misunderstanding,” where underspecified goals led to unsafe or misaligned outcomes, the pass rate was 0%.

Zero. Not low. Not disappointing. Zero.

My TenantKit experience isn’t an edge case. It’s the documented default.

The Showcase Illusion

If you visit OpenClaw’s official showcase, you’ll find a curated gallery of what people are building with the tool. It’s impressive at first glance. Look closer and a pattern emerges: booking a tennis court. Controlling a smart device. Automating a grocery order.

These are not applications. They are scripts with a conversational interface. They are short, linear, single-outcome tasks with a clear start and a clear finish. There is no ambiguity. There is no recovery from failure. There is no state to manage across days or weeks.

There are no verified examples in that showcase of OpenClaw successfully building and launching a complex production application from scratch. Not one. The demos that fuel the hype are demos of a very different and much simpler capability than what the hype implies.

This gap between demo and deployment is the central lie of the OpenClaw moment. The community celebrates booking a lobster dinner as proof that agents are building the future. They are not the same thing.

Confident, Not Cautious

The most dangerous thing about OpenClaw isn’t that it writes bad code. It’s that it writes bad strategy with complete confidence.

When I asked for the fastest path to revenue, I didn’t need a developer. I needed a strategist. OpenClaw gave me a developer, one that couldn’t even finish the job. It understood my words and ignored my context. It optimized for the literal prompt and missed the actual goal.

This is the failure mode that doesn’t show up in demos. In a demo, the goal is clear, the scope is contained, and success is easy to define. In the real world, goals are ambiguous, scope creeps, and the right answer is often “that’s not the right question.” OpenClaw cannot tell you that. It will just start building.

One security researcher put it plainly: “I ship code I never read.” That’s not a feature. That’s a confession. And it captures exactly what’s broken about treating OpenClaw as a co-founder rather than a very fast, very literal intern.

The Security Problem Nobody Wants to Talk About

The strategic failures are frustrating. The security failures are dangerous.

Fortune reported that giving OpenClaw the power to do useful things also gives it plenty of opportunity to do things it shouldn’t - leaking data, executing unintended commands, or being quietly hijacked by attackers. This isn’t a theoretical concern. Bitsight found over 42,000 OpenClaw instances exposed on the internet, with 93% of verified instances exhibiting critical vulnerabilities. Research has characterized the tool as “insecure by default,” with risks including plaintext credential storage.

The same trajectory-based safety audit found a 57% pass rate on prompt injection robustness, meaning nearly half the time, an adversarial prompt could redirect the agent’s behavior. You’re not just trusting the AI. You’re trusting everyone who might interact with it.

And the setup complexity isn’t getting better. OpenClaw’s own creator has since joined OpenAI. The person who understood the project best now has other priorities, and the onboarding reflects that.

Who Is This Actually For?

Here’s what I think is true: OpenClaw is a genuinely useful tool for a narrow set of tasks. If you need to automate a well-defined, repeatable workflow - something with clear inputs, clear outputs, and no ambiguity - it can do that well. The smart home integrations work. The booking automations work. The contained, linear tasks work.

What doesn’t work is the narrative that this is how software gets built now. That you can describe a business and wake up to a product. That the gap between idea and execution has been closed.

China’s tech sector is betting heavily on that narrative, with companies racing to deploy OpenClaw-powered products as a growth strategy. The enthusiasm is understandable. The foundation is shaky.

The emperor is wearing very impressive-looking clothes. They just don’t exist.

The Honest Assessment

OpenClaw passed Linux on GitHub because it represents something people desperately want to believe: that the hard parts of building software are solved. They are not. The hard parts - strategic judgment, error recovery, understanding intent over instruction, maintaining coherent state across a complex project - are exactly where OpenClaw fails.

58.9% safety compliance. 0% intent understanding. 42,000 exposed instances. A showcase full of booking demos.

This is not the future of software development. It’s a very fast, very confident, very brittle tool that is excellent at simple tasks and dangerous when given complex ones.

I wanted a business. I got a half-started codebase and a lesson about the difference between hype and capability.

The lesson was worth more than the code.

Have you actually shipped something meaningful with OpenClaw? I’d genuinely like to know, not the demos, not the automations. A real product, in production, built end-to-end. I’m skeptical it exists at scale. Prove me wrong.