Local models
Aiming to keep model services in your environment. Network access follows a separate policy.
Plan.
Execute.
Verify.
AOS aims to bring local models, tools and an isolated computer together in an agent environment with explicit state, outcome verification and human oversight.
FROM GOAL TO VERIFIED OUTCOME
Your request and boundaries
↓ Goal & scope
HOST
PlanBreak the goal into explicit steps
01↓ Proposal → policy check
ExecuteIsolated Agent Computer
02↓ Fresh observation
VerifyCompare the outcome with success criteria
03Explicit state · Persistent records · Human control
Explanatory diagram · not a running agent session
Explicit state
Tasks, actions and evidence outside chat: typed, persistent and traceable.
Human control
Explicit permission boundaries. Approval, rejection and handover are part of the flow.
01 / WHY AOS?Producing an answer
Producing an answer
is not completing a task.
For an agent, suggesting the right command is not enough. It needs to know what to do, where it may act and whether the result actually happened.
AOS separates these questions: planning proposes a route, policy checks authority, execution uses an allowed tool, and verification tests the result. “Operating system” here means this task environment; it is not a new kernel or Linux distribution.
02 / A CONCRETE EXAMPLEA small task.
A small task.
An explicit success criterion.
From creating a fictional file to verifying its contents. Explore the steps at your own pace.
Architecture explanation / example flow — not a real agent session.
Visible design.
Visible limits of evidence.
The 19 September 2026 package contains architecture and contracts. Actual models, target hardware, GUI runtime and training remain unverified. Open source is a design goal.
Explore status and acceptance criteria →