Observe
Read state and inspect the sandbox without ceremony.
03 / AI Laboratory · A1OS
Frozen as evidence · June 2026
Autonomy needs a handback.
A working agent operating-system prototype became the evidence for four supervision patterns.
The product was frozen. The design language continued: how much autonomy to grant, what to make legible, how to surface outcomes, and how to return control.
01
The consequence ladder
The next action decides
who should hold control.
Autonomy is not a single setting. As consequence rises, the system must reveal more intent, invite intervention and finally return agency explicitly.
Act visibly
Write or modify with a legible record of what changed.
Pause
Let the human approve, edit or reject with feedback.
Hand back
State what happened, what exists, where it is and what remains uncertain.
02
The operating cycle
Four patterns,
one supervision model.
The patterns were named after the prototype made their absence visible. Together they govern permission, explanation, outcome and control.
They are proposed interaction patterns—not standards or validated safety mechanisms.
Graduated Trust
Autonomy changes with the consequence of the next action.
Legible Autonomy
Plain-language intent first; raw telemetry remains available when needed.
Outcome Over Process
Deliverables are primary. The transcript is supporting evidence.
Accountable Handback
Completion is a designed return of control, not a buried final message.
03
The prototype evidence
Three situations
made the failures visible.
Execute, then ask.
A code task exposed the need to intercept consequential shell and file actions before they run.
Recover from error.
Autonomous recovery showed why a human needs a readable timeline rather than a wall of logs.
Reject and steer.
Human feedback had to become useful input so the agent could adapt instead of merely stopping.
Control never formally returned.
Outputs appeared, but ownership of the next decision remained ambiguous until handback became explicit.
04
The freeze decision
The prototype had proved the patterns—and buried them.
A1OS was compelling as a cockpit but contrived as a daily product. Its shared sandbox and developer framing made the supervision ideas vivid without making them broadly useful. The repository is frozen as evidence. The latest Studio handoff still lists the full pattern port as pending, so this page does not claim those patterns are already shipped there.