An open network for self-improving intelligence
Open-weight models that improve themselves from real use — served behind a standard API, owned by you, federated across nodes.
THIS TALK · THE OPEN FEDERATED PLAN 9 AGENTIC OS
No lock-in. Local, cloud, or both.
Run inference, training & Workers anywhere — everything is a file, so everything federates.
For this room: a Plan 9 agentic OS
The broad story is a platform for agentic infrastructure. The sharper lens — the one for this room — is an operating system for agents: everything is a file, per-process namespaces, capability-scoped by construction.
Everything is a file
A Plan 9-inspired virtual filesystem — models, streams, MCP tools, repos & sandboxed apps are all composable files in one namespace.
One abstraction, laptop → federation: the filesystem is the OS and the agentic interface.
Addressed by identity, not by host
Every node, model & stream is named by portable cryptographic identity — reached by who owns it, not where it's hosted.
Direct when you can, anycast when you can't — same identity either way.
The browser is a real node
A high-performance bus for real AI payloads — built to move tensors, tokens, video, and world-model frames between the model, your GPU, and your UI.
The browser computes and renders — a real participant, not a viewer.
A shell built for tools
Models are already fluent in shell syntax. HyprStream gives them a real one — Tcl (molt), exposed as files — so agents compose tools into programs instead of emitting one rigid JSON call at a time.
A language the model can write; types the machine can trust.
Agentic recursive self-improvement
Inputs → inference → outputs; the outcomes feed back, and the STEP loop turns them into a candidate model — driven by tool calls.
Stage and train speculatively · evaluate · then propose the winner.
Alignment is a merge request
A candidate isn't auto-promoted — it's proposed as a branch, evaluated, and reviewed by humans and agents. It merges only on consensus.
Consensus before promotion — every merge is reviewed, signed, and reversible.
Containers that react
Identity, separate from infrastructure
You keep your identity; workloads migrate across federated nodes — never tied to a provider.
Trust the keys, not the transport
Every STEP is observable
Traces, metrics & logs across STAGE · TRAIN · EVALUATE · PROMOTE — observe every stage of improvement.
One binary. Your hardware.
Universal AppImage or Nix flake, auto-detecting your accelerator · built on stable libtorch.
Tenstorrent — coming soon.
Bridged to the physical world
Real-world inputs in, real-world actions out — the same file interface, at the edge.
Your models train themselves. You version them with Git. You serve them with APIs you already know.
one open Rust binary · OpenAI-compatible API · MCP · Arrow Flight SQL
An open substrate — come build on it
Open weights, open protocols, verifiable by design. Run a node, extend the model, federate.
github.com/hyprstream/hyprstream
built on Wanix · ← all talks