This is a submission for the OpenClaw Writing Challenge
Context
This is going to be short post. The goal is for me is understanding Ope...
For further actions, you may consider blocking this person and/or reporting abuse
I wouldn't dream of using it personally. Not as a forever thing, but I have no interest in diving head first into this sort of thing too early on. But I see the appeal in general. For me I'm too old to be an early adopter.
AI agents by itself seems interesting to me and it's nice to try out! I just don't like the selling point where it can do anything. It just gives me anxiety to think about and it feels like it slows down production than anything since I have to make sure it understood the prompt I asked.
Thanks @ben! :D
Well, people gone so into "millionaire over night"-mode that AI decided to disable their tricky plans. My take on OpenClaw is that if one runs it on their private machine to run their private stuff, they just need to read my next article I will upload 😀😂
Make sense! I do despise the idea that you can make "millions" overnight just because an AI does a task faster and "knows" programming. Thanks Luftie!!
i never used openclaw, first because i never had the great setup to run it. there are so many good orchestrators out there u can build ur own openclaw. too muuch hyped
Make sense. I don't think you need a great setup for openclaw. You just need a good computer to run it, even locally. I see some impressive things being used for OpenClaw, but it's just not my cup of tea. Thanks Frank :)
Interesting take, but I think this is where it starts to feel a bit off.
I’m all for AI helping devs move faster — CLI tools, copilots, even things like Playwright MCP are genuinely useful. They still keep you in the loop. You’re driving, AI is assisting.
But tools like OpenClaw push closer to “just let the AI do everything,” and that’s where it loses me a bit. If you’re not thinking through the code, debugging, or making decisions… what’s your role anymore?
I get the argument for small teams — spinning up an “AI teammate” sounds great on paper. But in reality, it still feels early. Reliability, context, and trust just aren’t there yet (especially for anything beyond simple flows).
Also, maybe this is more of a personal take, but part of building software is… actually building it. Problem-solving is the fun part. If you remove that entirely, it kind of defeats the purpose.
Curious where others draw the line between “assistive” and “replacement.”
OpenClaw's whole point is replacing developers and laying off devs. It attracts product managers and non-tech people because it makes their dream come true where they don't have to hire any dev and give them monthly salary. Instead, the AI gents will do all the work and their company will be finally employee-free. No human in the loop, this is what makes it exciting for them.
Yeah indeed
Even that, AI in general is what makes people happy since they see AI as a replacement to humans when doing tasks, which I don't like at all. As a tool, make sense. I believe now, people are starting to realize that it is not a good replacement to human workers.
Great question! Depends on the individual of how they see OpenClaw. That leads to what you mention about the line between "assistive" and "replacement". It depends on one experience.
If a person uses it as a "replacement", they will eventually see reality that it is not the case down the line. It's a costly lesson, but I guess it's the only way for one to see the consequence since OpenClaw is more than just an "agent". It just makes completing tasks faster.
Thanks Mark :D
Powerful idea, but the blast radius is still too big.
Indeed.
The matching engine problem you hit is classic cross-platform entity resolution, and your 4-stage pipeline is actually pretty solid. A few thoughts from running into similar issues:
Your Entity Guard is doing a lot of heavy lifting. One thing that helped me was normalizing entity aliases before the Jaccard step â things like "NY Knicks" vs "New York Knicks" vs "Knicks" all mapping to a canonical form first. The Jaccard similarity between normalized strings is much more reliable.
Union-Find is the right call for transitive chains, but watch out for transitive false positives. If A matches B at 0.85 confidence and B matches C at 0.82, you're chaining AâC at effectively ~0.70. We added a max-chain-length limit of 3 to prevent this.
Have you considered using TF-IDF weighted cosine similarity instead of raw Jaccard? It handles the "France vs Brazil" vs "France vs Germany" edge case better because it weights discriminating terms higher.
The 95%â98% gap you're stuck on is brutal â in my experience that last 3% usually requires going from string similarity to actual semantic matching. Did you try any embedding-based approaches for the stragglers?
Great writeup either way. The architecture itself is clean enough that the matching layer is really the only thing holding it back.
Thanks :)
The part that resonates isn't the skepticism about OpenClaw specifically—it's the deeper unease about delegating agency to something that doesn't share your context or consequences. An AI that can "do anything" on your computer is fundamentally different from an AI that suggests code completions. The failure modes aren't just wrong answers; they're wrong actions.
What's interesting is that this hesitation exists even when the tool is open source and the code is auditable. Transparency helps with trust in the implementation, but it doesn't address trust in the execution. Knowing how the agent works doesn't tell you what it's going to do next Tuesday when you phrase a prompt slightly differently than you intended.
The token cost point is the practical anchor. People are stacking subscriptions—$20 for ChatGPT to run OpenClaw to run $200 worth of Claude Code. That's a real monthly bill for a workflow that's still experimental. It's easy to get swept up in the demo and forget that the meter is running the whole time.
Your FTX analogy is a stretch, but I think I understand the instinct behind it. It's not about fraud. It's about the gap between perceived safety and actual safety. FTX felt legitimate—mainstream sponsors, celebrity endorsements, a professional facade. OpenClaw feels safe because it's open source and well-documented. But the feeling of safety and the fact of safety are different things, and the difference only becomes visible after something goes wrong.
I'm in a similar camp of watching and waiting, but I wonder if there's a middle ground that doesn't require a dedicated Mac Mini. Could you run it in a Docker container with limited filesystem access? Or on a cloud VM you spin up only when you need it? Something that gives you the exploration without the commitment. Feels like there's space between "not yet" and "all in." Have you experimented with any sandboxed approaches, or is the whole category just on pause for you?
Hey! Thanks for taking the time to write your comment!
Pretty much. FTX does seem to be a stretch, but the goal is to understand that if it rises quickly, always be cautious about why. Even though it rises quickly into popularity doesn't mean it is "safe". It's just prone to hackers challenging themselves into finding vulnerabilities and other people looking deep into it faster. Obviously, it's not a "recipe to disaster", there are cases where one thing rises to popularity quickly and still ended up being just fine for a decade. It's just seeing that pattern makes me concern that it is the most likely case, but it is never 100%.
I have not, and probably never because of how scary it is for me. Theoretically, you could open up a Virtual Machine and run OpenClaw there, but the setup is interesting to where I don't bother. Maybe someone has done this before using a VM, but not sure.
Thanks for reading! Appreciated it :D
Searching job so i can get money to buy a computer just so i can run openclaw locally.
If its run with just tokens there is too big risk.
But local llm running OC if it needs cloud services etc i buy more hardware or use prepaid services so the error of OC wont break my retirement plans. If am worried about my personal data I would never have them in space where OC controls or has access to.
My assumption is we have to change how we operate to accommodate the openclaw.
That's fair. IMO, I wouldn't trust it now based on the pattern I see in products getting too much hype really quickly. Changing how we operate to accommodate OpenClaw is interesting, but it depends on if everyone collectively agrees for that change to happen in the first place. Thanks :)
Kinda agree with the caution tbh — giving an AI agent full “do whatever it takes” access to your machine before solid guardrails exist feels less like productivity and more like speedrunning a future postmortem 😅
"feels less like productivity and more like speedrunning a future postmortem" is one way to put it lol. My overthinking brain thinks that it's their plan to take over everyone's computer since it's an AI Agent that can do anything, which is a fair reason to be cautious. Thanks Willie :D
Interesting take — sometimes the best advice is to wait until promising tools mature before relying on them in production.
Indeed it is the best advice. It goes with anything too whenever a new product is release. It is better to wait until:
Thanks Laura :D
Worth separating two concerns here that often get collapsed into "is OpenClaw safe yet":
Capability risk — "it can do anything." This is a scope problem, and it's solvable today with run boundaries: disposable sandboxes, per-run credential scoping, denylist the file system and network by default, require explicit grants for side effects. If a tool doesn't give you those knobs, the problem isn't "wait a year," it's "this tool isn't shipped."
Trust-the-vendor risk — new project, rapid growth, possible supply-chain or governance failure. This one actually does get better with time, and the FTX analogy fits. But I'd frame it as "run it behind your own auth/audit layer," not "don't run it."
Token burn is the one nobody can design around — agent loops are structurally expensive and will stay expensive until model prices drop another order of magnitude. That's the real reason most teams I've seen are holding off, not the safety story.
Thanks for your comment!
This can be interchangeable. If someone were to say "wait a bit more" they meant that the tool isn't ready. Additionally, it depends on the person as well. For example, one of my friends decide to not buy a VR headset. Not because it is "not ready", it's because he wanted to wait to see the better version of it. Many people have different perspectives on if the tool is "ready" or not.
Token burning is a big factor and I can agree on that. I know there are solutions to this to reduce burning down tokens, especially LLMs in particular. Though, I don't know on the top of my head but I know there are solutions out there to minimize token usage.
Thanks Aikit!! :D
I absolutely love the raw honesty of "I wrote a post about why you shouldn't use this tool that I have personally never used." That is peak developer blog energy and I am 100% here for it. 😂
You raise some totally valid points though! Handing an AI the keys to your entire operating system and saying "do things" is essentially the plot of every sci-fi thriller from the last thirty years. It's all fun and games until OpenClaw decides the most efficient way to clear your inbox is to format your hard drive. And the token burn... ouch. Imagine waking up to a $40,000 API bill, because your AI agent spent the night arguing with another AI agent in a Discord server about the optimal way to write a regex.
I'll probably still end up trying it out on an old, sandboxed Raspberry Pi that I keep in a Faraday cage...or, my old "gently used" Mac Mini, but I appreciate the warning! 🦀
Hey Cyber! Great point where you mention "It's all fun and games until OpenClaw decides the most efficient way to clear your inbox is to format your hard drive.". It's a scary thing to think about lol.
Feel free to try OpenClaw on your Raspberry Pi! Make sure it doesn't "break out" somehow! Thanks for reading :D
stability isn't the actual problem. debugging when an agent gets stuck in a loop - that's what's rough right now. the core orchestration works fine if you know what you're doing.
I see! Makes sense for that to happen. Pretty much the whole idea is OpenClaw in general since it rose up quickly, so it's a good idea to be skeptical at the moment. It's not about how stable it is right now, it is about the pattern we see in products in general. Hope this makes sense! Thanks for reading Mykola :D
the pattern read makes sense. most new orchestration tools hit this wall - not whether it works, but whether the primitives hold under real load. that's the actual question.
been rocking hermes and nanoclaw. both simpler and (i think) more secure.
🤔
Interesting perspective on OpenClaw! Would love to see more details about your setup 🔧
Thanks Clavis! Could you clarify on what you mean by "setup"? I mention this because I did stated that I am not planning on using OpenClaw. Regardless, thanks for reading :D