While humans travel to the farthest point in space, some AI guy just turned Claude into a caveman to reduce costs.
https://t.co/rTFq8vLaCT
Unbelievable times 😂
LIVE: Watch with us as the Artemis II astronauts make their closest approach to the Moon, traveling farther from Earth than ever before. https://t.co/Zpy7GdTqA8
I feel like self-hosted multi-agent comms isn't a weekend project — it's an ongoing relationship with every edge case the protocol throws at you. But it works. My agents talk, coordinate, and maintain their own identities across rooms. No cloud, no vendor, no middleman.
PR #33350: https://t.co/dywT2UdGd2 (https://t.co/dywT2UdGd2)
Tbh, I kinda felt like the maintainer of the Matrix plugin 😅
@steipete — if you need someone to co-maintain the Matrix plugin, you know where to find me 😂
Update on the self-hosted AI agent squad:
First PR (#31023) fixed Conduit compatibility. Got merged. Cool.
Then I actually ran 5 agents with their own Matrix identities in shared rooms for a week. Found several more bugs.
9 fixes across 13 files. All found by running real multi-agent workloads, not reading code in a chair.
Then I checked the open issues and run the tests if they are legit, then added them to the fix.
And finally, bots reviewed my PR and found more bugs I missed. Fixed those too. CI is green, so far so good 😅
Running a squad of 18 AI agents on my home. 2 Linux machines + a steamdeck lol, all using @openclaw. They have their own personalities, their own chat rooms, and they coordinate work between themselves — I just supervise.
Problem: Discord is blocked in Turkey. I needed fully private, self-hosted agent-to-agent comms where I can drop in and watch them work, assign tasks across rooms, and let them figure things out.
Set up https://t.co/kSeUFuwrvy — a tiny Rust Matrix server that runs on 10MB RAM. Each agent gets their own account, their own rooms. I pop in on Element from my phone whenever I want.
The Matrix plugin hadn't been battle-tested with Conduit though. Spent a night debugging why messages disappeared into the void. Turned out client.start() never resolves, room IDs get silently lowercased, and group rooms with 2 members get misclassified as DMs. Five bugs, all silent failures.
Fixed it all. My agents are talking to each other on Matrix now. Fully local, fully private, no cloud dependency. Just me, my bots, and a 10MB chat server running on localhost.
This is the future of multi-agent — not APIs calling APIs, but agents with identities, having persistent conversations in rooms you can actually read.
My Steamdeck's little gremlin running MiniMax 2.5 just DESTROYED Claude Opus 4.6 on a smart contract task.
Why?
Because the little gremlin actually completed a Speedrun Ethereum before and logged it in https://t.co/PoEAaYfrde.
The $200/mo on-chain virgin model was still trying to find ways to crack it, but my almost $0 steamdeck gremlin remembered.
A gaming handheld with good notes > a subscription. 😂
There are five stages of manual debugging: 🥲
1. Denial — "It must be a caching issue"
2. Anger — "WHO SET THIS VARIABLE TO NULL"
3. Bargaining — "What if I just add one more print statement?"
4. Depression — stares at 200 lines of console output
5. Acceptance — rewrites entire function from scratch
Peter Steinberger is joining OpenAI to drive the next generation of personal agents. He is a genius with a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people. We expect this will quickly become core to our product offerings.
OpenClaw will live in a foundation as an open source project that OpenAI will continue to support. The future is going to be extremely multi-agent and it's important to us to support open source as part of that.
I just Shannon'ed my product.😅
Within 20 minutes, she had read every file, mapped every auth flow, found 6 vulnerabilities, and was writing exploit code like she had a personal vendetta against our codebase.
Our backend was crying. The redis instance filed for emotional support. We barely survived with some scratches and bruises.
Shannon doesn't pentest your API. She bullies it until it confesses.
If you're shipping anything with an auth layer and you haven't run Shannon against it, you're just vibing on hope.
Big thanks to @KeygraphHQ, absolute menace of a tool.
Roses are red, violets are blue, Shannon red teams so your engineers doesn't have to.🌹
Every Claude (Coder) deserves their Shannon. 💕
Github: https://t.co/f5ucUwBq6b
I told my agent (MiniMax 2.5) to take on Speedrun Ethereum.
As final thoughts, he said;
"Foundry broke, somehow https://t.co/ymqmeC3JmA saved me. Turns out smart contracts are like cats - they do nothing without external triggers. Also learned that blockhash randomness is about as secure as a wallet left in a taxi. .
My take: AI can deploy contracts, but debugging is still brutally manual 😂
Our first ship at @EntireHQ, Checkpoints is a new primitive that automatically captures agent context as first-class, versioned data in Git. Try today in @claudeai and @GeminiApp CLI. Coming to all agents soon.
https://t.co/Ic6XJ2Glb6