> just like LLM agents were a new layer on top of LLMs, Claws are now a new layer on top of LLM agents, taking the orchestration, scheduling, context, tool calls and a kind of persistence to a next level.
Layers of "I have no idea what the machine is doing" on top of other layers of "I have no idea what the machine is doing". This will end well...
Yeah, in the interest of full disclosure, while Claws seem like a fun toy to me, I tried ZeroClaw out and it was... kind of awful. There's no ability to see what tools agents are running, and what the results of those tools are, or cancel actions, or anything, and tools fail often enough (if you're trying to mind security to at least some degree) that the things just hallucinate wildly and don't do anything useful.
The ZeroClaw team is focusing their efforts on correctness and security by design. Observability is not yet there but the project is moving very rapidly. Their approach, I believe, is right for the long term.
There's a reason I chose ZC to try first! Out of all of them, it does seem to be the best. I'm just not sure that claws, as an overall thing, are useful yet. at least with any model less capable than Opus 4.6 — and if you're using opus, then whew, that's expensive and wasteful.
The ZC PR experience is hard core. Their PR template asks for a lot of details related to security and correctness - and they check it all before merging. I submitted a convenience script that gets ZC rolling in a container with one line. Proud of that!
Regarding models, I’ve found that going with OpenRouter’s `auto` model works well enough, choosing the powerful models when they seem to be needed, and falling back on cheaper ones for other queries. But, it’s still expensive…
Depending on what you want your claw to do, Gemini Flash can get you pretty far for pennies.
> Layers of "I have no idea what the machine is doing" on top of other layers of "I have no idea what the machine is doing". This will end well...
I mean we're on layer ~10 or something already right? What's the harm with one or two more layers? It's not the typical JavaScript developer understands all layers down to what the hardware is doing anyways.
I will assume you know that comparison is apples and oranges. If you don’t, I’d be happy to explain.