Octopus SI OS Rebuilds the AI Agent Runtime as an Open-Source Rust Kernel
A Rust-based agentic operating system where agents rewrite their own harness mid-task, built atop Codex with a 45k-line kernel.
- Octopus SI-OS is an open-source agentic OS with a 45k-line Rust kernel, MIT licensed.
- Agents can edit their own harness mid-task by writing JSON and small scripts.
- Built on top of the Codex source, with a macOS 14+ desktop client and voice mode.
- No central orchestrator: primitives are Interfaces, Artifacts, Runs, Environments, Workers, Policies, Marks.
- Deny-by-default sandbox, local-only data in
~/.octopus, every action logged in a ledger. - Research preview: author admits current models, including GPT-6, do not yet exploit the self-modification surface.
Octopus SI OS is an MIT-licensed research preview that recasts an AI-agent harness as a local, operating-system-style runtime. Its roughly 45,000-line Rust kernel manages durable runs, policies, interfaces, artifacts, and verification, while models and tools connect through those primitives. Here, “OS” refers to the runtime architecture; the software runs as a macOS application.
By placing the runtime at the center, Octopus lets agents create tools, register interfaces, and modify their working environment while a task is running. The first packaged interface is a desktop client derived from Codex, with ChatGPT sign-in and Codex CLI available inside the broader system. The project explores whether persistent state, scoped permissions, and live extension can help agents operate beyond a single model-and-tool loop.
Inside the runtime
The desktop client borrows its interaction model from the ChatGPT app and organizes agent work around five OS-inspired concepts:
| Concept | Role |
|---|---|
| Home | Maintains a continuous text or voice conversation and renders live widgets while work proceeds. |
| Projects | Isolate goals, context, and permissions so capabilities remain scoped to the relevant work. |
| Agents | Represent workers through two JSON files, allowing multiple agent definitions without separate programs. |
| Tasks | Store durable units of work that can pause and resume later. |
| Drivers | Connect actions to folders, Git worktrees, browsers, desktops, containers, and remote machines. |
The ledger drives the kernel
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
