this was all made in ~5 days of constant vibe-coding using MiMo V2 Pro (on medium and high, though usually medium) in opencode. 0 lines of human-written code.
000 - what
if you have a job, a social life, any remaining will to live, or any combination of those:
- an ai agent/harness is something that allows an llm to have (sometimes unrestricted) access to an environment
- these are usually split into 2 categories, ones for coding (claude code, codex, cursor, opencode, and the like), and so-called "claws" inspired by @steipete's
clawdbotmoltbotopenclaw (which is by itself a horrible project but that's off-topic) this project is a claw
001 - claws
there are 2 competitors in the claw space right now:
- openclaw by peter
- hermes-agent by the nous research team
openclaw came first and came extremely flawed from the start, with complete disregard of code quality, security, stability, and usability. after it came various projects that never reached an audience. and then came hermes-agent. it's pretty good, a lot better than openclaw, very local ai friendly. i ran it for a little while but didn't achieve much with it, it felt off.
010 - lume
lume is "my" personal attempt at making a claw to fit my style of doing things. "my" is in quotation marks because it's entirely vibe-coded with 0 human fixing and near-0 review, just to see if it can go as badly as openclaw. it really didn't go all that badly.
011 - what works
it's not secure - intentionally. the ai tried to convince me to add sandboxing and tool checking a million times, but i declined.
for the agent itself, i had 2 of my friends mess with it for hours and all they managed to do was convince it that it's a (this is a direct quote from its SOUL.md) "sassy, confident Asian baddie". as for capabilities, they got it to search the web 3 times, once for what trade to make on nasdaq - successfully (both in searching the web and in not longing right before a crash), once for recent news - successfully, and once for "shoujo tsubaki" - unsuccessfully, stuck in a loop. 2/3 is a bad track record though probably an issue with the toolcall.
100 - what doesn't work
the things that i couldn't get working at all:
- cron jobs - the agent can work with them fine but i couldn't get ai to fix the scheduler in the harness
- the system prompt - the agent doesn't follow it half the time
- skills and plugins - these could probably work if i had a way to give the agent access to documentation but i haven't been able to think of a way to do it without bloating context
the things that sort-of work:
- the memory system - the model uses it sometimes, though i don't think the way it's injected into context is good
- the computer use - it can use the tools to some degree, though i think i need to change them to be similar to mainstream harnesses to get the model to prefer them over bash
101 - discord interaction
one of my big goals with this was to make it work well in discord, well enough for it to be the main channel of interaction. i feel like i somewhat succeeded in this, from testing it handles small groups of people (2-4) talking to it at once pretty well, context doesn't lose itself, messages mid-generating act as steering (though there are some tiny edge-cases with this which need to be fixed), and with a good SOUL.md it feels almost natural to talk to.
a lot of things still need working on, but i am somewhat confident that this can actually work and feel good to use.
110 - self-alteration and self-improvement
these are, at least from what i've seen, two of the biggest things that people look for in agents - the agent's ability to grow with the user. it was also my other big goal.
along with memory and skills, i wanted to give the agent the ability to write its own code via plugins, in the form of rhai scripts. as of right now, none of that works. hell, even skills don't work. i bit off a lot more than i can chew.
111 - where to from here
for the ai industry: same place they were always going.
for me and lume:
- i made the lume repo public (funnily enough, on a broken commit) here, none of the code except the README.md has been touched by me at all
- i will continue prompting the ai to work on it, maybe i'll even work on it myself a tiny bit
- if anyone feels the urge to use it, there's no prebuilt binaries but it's as easy as
cargo build --releaseand then runninglume daemon, which should generate the default config in~/.config/lume/lume.toml - if anyone feels the urge to contribute, i do check issues and PRs on it
this post is meant more as a dev-log for what i've spent the past week doing, but if you for whatever reason decide to try this at all, do tell me about it, you can find where to contact me on this site's home page.