I've been messing with it the past couple days. I put it in a VM, on an untrusted subnet I keep around for agentic stuff. I see promise, but I'm not especially impressed right now.
1) Installation on a clean Ubuntu 24.04 system was messy. I eventually had codex do it for me.
2) It has a bunch of skills that come packaged with it. The ones I've tried do not work all that well.
3) It murdered my codex quota trying to chase down a bug that resulted from all the renames -- this project has renamed itself twice this week, and every time it does, I assume the refactoring work is LLM-driven. It still winds up looking for CLAWDBOT_* envvars when they're actually being set as OPENCLAW_*, or looking in ~/moltbot/ when actually the files are still in ~/clawdbot.
4) Background agents are cool but sometimes it really doesn't use them when it should, despite me strongly encouraging it to do so. When the main agent works on something, your chat is blocked, so you have no idea what's going on or if it died.
5) And sometimes it DOES die, because you hit a ratelimit or quota limit, or because the software is actually pretty janky.
6) The control panel is a mess. The CLI has a zillion confusing options. It feels like the design and implementation are riddled with vibetumors.
7) It actively lies to me about clearing its context window. This gets expensive fast when dealing with high-end models. (Expensive by my standards anyway. I keep seeing these people saying they're spending $1000s a month on LLM tokens :O)
8) I am NOT impressed with Kimi-K2.5 on this thing. It keeps hanging on tool use -- it hallucinates commands and gets syntax wrong very frequently, and this causes the process to outright hang.
9) I'm also not impressed with doing research on it. It gets confused easily, and it can't really stick to a coherent organizational strategy over iterations.
10) also, it gets stuck and just hangs sometimes. If I ask it what it's doing, it really thinks it is doing something -- but I look at the API console and see it isn't making any LLM requests.
I'm having it do some stuff for me right now. In principle, I like that I can have a chat window where I can tell an AI to do pretty unstructured tasks. I like the idea of it maintaining context over multiple sessions and adapting to some of my expectations and habits. I guess mostly, I'm looking at it like:
1) the chat metaphor gave me a convenient interface to do big-picture interactions with an LLM from anywhere;
2) the terminal agents gave the LLMs rich local tool and data use, so I could turn them loose on projects;
3) this feels like it's giving me a chat metaphor, in a real chat app, with the ability for it to asynchronously check on stuff, and use local stuff.
I think that's pretty neat and the way this should go. I think this project is WAY too move-fast-and-break-things. It seems like it started as a lark, got unexpected fame, attracted a lot of the wrong kinds of attention, and I think it'll be tough for it to turn into something mature. More likely, I think this is a good icebreaker for an important conversation about what the primetime version of this looks like.
I have set it up in a docker container that only has two host volumes mounted with the Obsidian vaults where I keep notes for two TTRPG campaigns, so a very low-stakes situation. I have set it up with a Discord bot so my players can chat about campaign and rules stuff (I already had player-facing notes for every session in the Obsidian vault in chronological order, plus a bunch of Markdown files with rules).
If the agent goes rogue and nukes my Obsidian vaults I have them backed up to Github private repos anyway which the agent cannot touch because I am not crazy to give it my SSH credentials.
I initially tried using Kimi-2.5 through OpenRouter and had the same experience you had with pretty bad tool use, not sure if this is a provider issue since it is a pretty new model. I switched to Gemini 3 through the Google AI Pro account I have for personal use and it was a lot smoother after that.
I have some experience with coding agents using Cursor for work and Antigravity for personal stuff, and the OpenClaw harness definitely seems worse, but for my low-stakes use-case I managed to paper it over with some edits to the AGENTS.md file.
But even in this very crude state it was already interesting to see one of my players giving the agent some info about his character, including an avatar image, and have the agent create a folder in my Obsidian vault to store this and update its memory file to be able to remember it for future interactions.
Scott Adams shaped my sense of humor and perspective on a lot of things. Even in later years, when I disagreed with him immensely on a lot of things, I found that there was a thread of insight in what he said regarding how people experience reality and the power of words and images. Ultimately I tuned out, but before I did I followed his line of inspiration (which he was very public about, often naming books and authors) for a lot of that and was not disappointed. I was grateful that the insight was again sincere, and learning them didn’t take me to the places I did not want to go — the places he himself seemed to sincerely enjoy.
It’s not hard for a lot of us to criticize who he became. He certainly had no shortage of criticism for others. I looked up to Scott a lot as a kid, and as an adult found him to be a man like any other, with limits and flaws… not merely in spite of his accomplishments, but often because of them. There’s a lesson there that I wish to carry too.
Whenever this stuff comes up I try to remind people that we only get to see a tiny little glimpse of what these folks were up to. Folks look at the stone tools that have only been found after their owners were done with them and left them in the ground for eons, and imagine that in general everything was “rough” and “crude.”
We don’t get to see the overwhelming majority of their craft — there’s no doubt a whole world of wood and leather artistry and so on that don’t get to survive. Humans are clever, adaptable and often times really fucking obsessive. That same instinct that makes one spend hundreds of hours on Factorio was around in prehistory, applying itself to whatever.
I often times hear anthropologists speculate that large stone handaxes were a means of seduction — that the girls would have swooned over the guys who were better able to make more effective tools. I know too many nerds to believe this. I think that back then, there were people who kept obsessing over making finer tools and theorizing about designs and where materials could be found, and it was about as sexually appealing as my homelab. Which is to say, absolutely fucking not, but who cares, I want to tell you about my idea for a subnet optimized to allow doomcoder agents to handle their own infra needs
Having a great hand axe might be more analogous to the boy with the “bitchin Camaro” to attract the girls. Or more recently, performance in a street takeover.
It’s really a shame how popular it was to mar shows with this… I saw a DVD set of a show once with a no-laugh-track version. It sucked because the actors pause for the laughs after each line. This is bad enough with the laugh track in place, but if it’s just dead air it makes every scene feel awkward.
for 1), seems like you could do a proxy encryption solution.
edit: wrong way to phrase I think. What I mean to say is, have a message key to encrypt the body, but then rotate that when team membership changes, and "let them know" by updating a header that has the new message key encrypted using a key derived using each current member's public key.
The obvious error here is that an inner voice is not fundamental, and the fact that many people describe their consciousness in such different terms makes it much more likely that consciousness is just something that manifests in a variety of subjective experiences.
I’ve had some limited success attributing ideas to other people and asking it to help me assess the quality of the idea. Only limited success though. It’s still a fucking LLM.
reply