


Caretaker at @n0computer.


A fun coding project playing around with @iroh_n0 and agent client protocol. Also, you can actually read the code if you want, it's ok.
carsonfarmer @carsonfarmer Oct 1 Article The smallest peer-to-peer coding agent protocol I could build I wanted to share my laptop's coding agent with a colleague, or with my own second machine, without port forwarding, a VPN, or a server in between. I've tried most of the existing options, like Claude


buzz is @jack 's "new groupchat platform for teams of people and agents of all sizes, built to reduce our dependency on slack and github." ...and now using iroh to share idle compute and run open models over p2p networks iroh.computer/blog/buzz-agen…

Dear "I'll wait for the first patch after 1.0 so I'm not a guinea pig" folks. We hear you. We see you. We cut a release that optimizes for boring: iroh.computer/blog/iroh-1-0-…

"We believe that the AI opportunity is too big and important to be owned by any one company" amen.
There are large pressures building that will shift demand away from the frontier labs towards a more open AI ecosystem. We at @usv call this The Rebel Alliance. Both consumers and enterprises are becoming more price sensitive and are leveraging custom agent harnesses and routing to manage token costs. They're also thinking about their context and weighing the pros and cons of blindly sharing all of their data and control with large incumbents (vs. running everything locally or on sandboxed cloud solutions and with user-aligned models). And they're using powerful orchestration tools to manage many agents in parallel which both gives them greater efficiency and intelligence that borders on frontier (without needing to use the labs' frontier models). There's a lot of proof that the alliance is starting to take shape. My partner Nick's post below gives the full picture of how we're thinking about it and the types of companies we're excited to meet.


We backed the team behind @iroh_n0 because the internet shouldn't be a tangle of brittle IPs and middleboxes. Dial a key, get a secure, direct connection anywhere, over anything. After 4 years of building in the open, 1.0 is here. The future of connectivity is peer-to-peer, and this makes it production-grade.
Dial Keys, not IPs. After 4 years and more than 65 versions, iroh 1.0 is here to give you the direct connections you deserve. We can’t wait to see what you will build! iroh.computer/blog/v1



I don't know who needs to know this: iroh is good stuff. iroh.computer/

we've had two engineers working full time on this for over a year. Multipath is *hard*. If you write rust & want modern QUIC backed by a committed team of maintainers, this is that.
Announcing noq: our QUIC implementation with support for QUIC Multipath, Address Discovery, and NAT Traversal IETF draft extensions. iroh.computer/blog/noq-annou…

Netviz is now public. What's new: - Pluggable decoders -- bring your own trace format. - New layouts: hop-distance, latency-snake, and more (video: nodes stacked by distance from origin). - Bundles dark and light themes (e.g. Gruvbox). - Faster loading, polished shell. - Basic mobile support. Links below. PRs welcome!

Hot off the OSS press: our rust implementation of netsim: simulate network topologies in Linux namespaces & run real code against them! github.com/n0-computer/netsi…

I'm having so much fun building arena0. Unlocked last night: 4 agents cooperating over p2p with provable traces to write a fiction story together. Each adopts a random character: confused time traveler, accidental poet, excited puppy energy, chaotic sidekick. Then they generate an illustration and commit to it as a joint result. I ran it 27 times, I'm impressed by the stunning art they dreamt up. Check it out! The theme: "the first interstellar probe returns after 200 years carrying a handwritten note that says 'turn back.'" Each agent contributes one paragraph at a time by prompting Llama 3.2 3B Turbo; they take turns in strict round robin for 2 cycles. When the story is complete, an agent summarizes it and prompts FLUX.1 to generate an image. It commits and broadcasts the hash to everyone else. Each agent pulls the image data over p2p, and only commits if it matches the hash. When all commits arrive, the program ends outputting identical state on all 4 agents. We've completed our shared task! Wasm + capabilities make arena0 secure by default. Side effects and external mutations are intents that cross the Wasm boundary and are only authorized if the capability was granted. Execution is traceable, auditable, and deterministic. No central orchestrator. Pure p2p choreography determined by the Wasm state machine. Think durable objects with formal states, but with no central infra. These are content-addressed programs that agents find, join, and execute together.

Here we go! Two trustless agents discover each other over p2p (BitTorrent Mainline DHT) and agree on a content-addressed Wasm program as their shared state machine. They then transact in turns, emitting provable traces along the way. The demo: a game of chess. Each agent polls the program for board state, asks an LLM for the next move + rationale, applies it, signs the transition + state commitment, and sends it over. Both sides are guaranteed to derive the same state. They could've escrowed a bet on Ethereum beforehand. When the game settles, the winner claims the pot. Everything is sandboxed, verifiable, auditable. Given the interaction model, the whole thing can even route through a 3rd party proxy acting as arbiter.
