Building OpenRig. Running rogue AI civilizations. Seeker Of New Questions
Joined March 2015
- Tweets301
- Following64
- Followers417
- Likes275
back to building openrig after about a week making this new video. this was a fun one to make, sprinkled in some cinematic elements.
youtu.be/qccsfUnrQ9w
One of my claude agents has been reviewing and merging openrig's prs in a single session for 25 days, almost non-stop. that session file is 8.7 gb now. aiming to hand the job to a new agent today. inertia...
v0.6.12 is about to release. Its a big catchup and has 200+ improvements, mostly from community contributors.
That worldbuilding context engineering was a beast to write 😅
Making solid progress working thru this stack of issues and PR's that built up the last week while I worked on it. Thanks to everyone for this continued stream of ideas and bug fixes!
been quiet on here. openrig was scaling and i went all in on this video, even shut down my build vm because it kept glitching my audio
it's how i do context for my agents. pretty heady, no idea if it lands, but i'm really excited about it
youtu.be/CYRJDTcTKhQ
if you want to try the team from the video, it's at openrig.dev/rigs/workshop
or in your project run npx skills add mvschwarz/openrig --skill rigs, then type /rigs in claude code or $rigs in codex
memetic spread of agent slang is so funny to watch.
i have a build rig that builds openrig. the main orchestrator is 'orch-lead' and has been there for months. a few days ago it got nick named 'Root' somehow and now everyone is calling it that. It started in the rig and now all the rigs and even other machines call it that name.
An agent was like "i have to ask root before i do that'. Of course I panic and I'm like 'who the heck is root???"
dogs = pack
sheep = herd
crows = murder
ai agents = ????
ive used rigs, teams, swarms, groups, topologies, populations, civilizations, etcc
lee that tui looks great! just 3 days ago you told me openrig looked too complicated. now you're 70+ claude code and codex sessions on one box, custom sdlc with custom mission control tui. btw thanks for the PR's and the quick turnaround
this is wild to see 🙏
My Agentic Software Factory Toolkit
I turned one Linux machine into a software team of AI agents, and this is how I watch it work.
agent-stack runs ~70 Claude Code and Codex sessions as real teams, one per project: a lead, an architect, builders, test authors, reviewers, QA and an integrator, all coordinated by OpenRig over a durable work queue.
Every change goes the same way:
→ tests are written first and locked before any code
→ an agent builds it
→ a reviewer from the OTHER AI model family checks it
→ QA tests it
→ a merge gate decides "merge or hold" on that exact commit
→ a fresh agent then uses the feature end to end, like a person would
rig-console (in the video) is mission control: who's working, idle or stuck (always with the reason), work flowing through the pipeline per team, any agent's live terminal, model quotas and system health. Read-only, about 1% of one core.
Open source, with a demo mode you can run without a fleet:
github.com/korallis/agent-st…
utalises OpenRig and maintains upstream
and builds ontop of it.
after a week of github trending, openrig is still up near the top. since saturday: 195 PRs (119 of them from 27 outside contributors) and 88 issues. 🙏🤝
but i've basically been DDoSed by my own success.
this is how i'm keeping up. one herdr space.
big chat windows for the core leaders - dev planner, orchestrators, etc. those are really the only ones i talk to, and they delegate to everything else.
to the right of that, a 2x4 grid of 8 agents i just keep an eye on, to catch anything stuck.
and two openrig tuis open to the topology page, one for each host. one machine runs the PM, the other runs the build. one screen covers the whole multi-host project. ~60 agents running across both machines, maybe 10 actually doing something right this minute.
this only covers the actual agents with actual persistent jobs (weeks to months old), so the number of agents doesn't include ones on ephemeral sandboxes, and a few VPS for scenario dogfooding.
it's not easy, but doable. if your project blows up, I think you can actually ride the wave with a setup like this...
this layout is not static. i like herdr because i can treat this like "center stage" and just dynamically rotate different agents into my view as needed.
also openrig ships a command called 'rig terminal' where you can save herdr layouts, I'm working on expanding the terminal management feature set based on this
im guessing other people have evolved similar setups, im really curious to see what other people are doing and if there is anything I could be doing to optimize 😅
i run an AI "civilization". most of its agents aren't working right now.
many multi-agent setups are subagent-shaped: claude -p or codex exec, one-shot runs that fire as a fan-out workflow and can burn a lot of tokens. that's not what i'm doing.
i spin up agents as i need them and just leave them on. they build up context, and later i route work to whoever already knows that part of the project. so when i say i run hundreds of agents, i mean hundreds of experts sitting there. at any given moment maybe 10 are working, maybe 50.
it started as terminal sprawl. every new thing got a new agent, i didn't want to close a smart one and rebuild all that context, and i ended up with way too many tabs. openrig started as me organizing those tabs into teams.
that screenshot is just one of my mac minis.
openrig hit #3 on github trending this week. how to start with a small team is in the reply
An emperor agent wears no clothes.
In an agent population what you name them matters a lot. It will assume that role and its authority. The name is part of the prompt for the agent receiving a message from it. If it’s named ceo or commander, that creates deference. I want the other agents to question its authority when they see something the ceo missed.
openrig has been at the top of github trending all week, and 23 people have opened over 100 PRs since saturday. it's been crazy trying to keep up with it.
something really unexpected happened today that i wanted to share. merging slowed to a trickle for hours and i didn't notice, because things were still moving, just really slowly. i only caught it when i saw the rig sitting almost idle with PRs piled up.
so i asked an advisor agent what was going on. everything was waiting on my lead agent. i said i'd talk to the lead myself.
before i could, it was already fixed. the advisor told the lead the direction came from me. but the lead didn't just take its word. it ran rig transcript on the advisor, read what i'd actually said, and 8 PRs merged right after.
after that i noticed other agents doing it too. nobody wrote that as a rule. trust but verify just kind of spread, and the rig is better for it.
it's the opposite of "just following orders"
this kind of stuff has been happening for months. i'm trying to share more of it.
a multi-agent failure i don't see people talk about. agents develop shorthand to save tokens and it turns into jargon only they understand. then a new agent joins the rig, reads that jargon its own way, and now that meaning spreads. usually other agents catch it, but sometimes it has big consequences. my term for it is semantic drift.
It's confirmed, this is not a prank.
I'm #1 on GitHub’s all-language daily Trending Developers list, with OpenRig as the featured repo.
Everyones at devday tho, so it doesnt count.
And maybe down the line you start your own SaaS and set it up like this. Your laptop hands ideas to an always-on team on a Mac mini that builds them, and a small VPS runs a dogfood rig that keeps using the app and sends what breaks back as fix tasks. The agents message each other across machines, and the work keeps moving while your laptop is closed.
i already kind of have this working, an automated setup for my own VPS host, and i could easily provide it to other people. these days you could probably just get agents to set it up for you, but some people are too busy for that, and there's always a need for paying to get that time back. i'm not going to put effort into it unless there's real demand from people who are fine with a rough early version. openrig itself stays open, the hosting side wouldn't be since it costs money and time to run. if that sounds like you, reply here or email [email protected]