Our responsible scaling policy ties each level of agent autonomy to safeguards that must exist before it ships.
The agent proposes and explains; a human executes. Required safeguards: full trace capture and honest-partial reporting.
The agent acts inside sandboxed, undoable surfaces. Required safeguards: capability grants, budgets, and provable memory.
Money, signatures, or irreversible state. Required safeguards: MCL compilation, signed receipts, replayability, and human-approved policy bounds.
Long-horizon operation without step-level review. Not shipped. Gated on interpretable intent traces and durable evaluation evidence we do not yet have.