An OS where the agent can rewrite its own house rules

What it does

Octopus is a Rust research OS for running agents: they read and write files, drive browsers, use your desktop, and you talk to them by text or voice. The twist is architectural — a ~45,000-line kernel where capabilities are contracts, not plugins bolted onto a loop. Codex CLI runs as one interface of the system (sign in with ChatGPT), and everything else — providers, channels, sensors, tools — attaches as a folder of JSON plus an optional script, without touching the kernel.

The interesting bit

Self-evolution here means an agent can, under policy, create goals, add tools, change its triggers, and register new interfaces — and those changes persist across runs, shaping what it notices and can do next. The README is refreshingly honest: today’s models don’t take initiative on their own, and for ordinary coding tasks Octopus gives you nothing Codex CLI doesn’t. The actual research question is whether an environment agents can reshape leads to real autonomy.

Key highlights

• Agents are two JSON files, not programs — spin up as many as you like, each with a role and a hard ceiling on what it may touch
• Every action is a row in a ledger; refusals are recorded too, never silently dropped
• Tools run in a deny-by-default sandbox; everything stays local in ~/.octopus
• The system can build its own missing pieces — e.g., author a camera driver and register it as a new interface, live
• macOS 14+ desktop client with voice, plus a terminal client sharing the same state

Caveats

• Research preview: stored formats may change between builds, and the packaged app is ad-hoc signed, not notarized
• macOS-only for the desktop client; requires a ChatGPT account
• By its own admission, no practical gain over Codex CLI for normal coding work yet

Verdict

Worth a look if you’re interested in agent architecture as a research direction — the ledger-and-contracts design is genuinely novel. Skip it if you just want a coding agent that ships features today.