The Autonomy Ladder

ContextStack separates permission to work from permission to create a lasting consequence.

Run-autonomy lets an accountable Bot commission bounded execution without an operator watching every step. The run still has a budget, capability envelope, external acceptance, and a recorded terminal state.

Consequence authorization governs merge, deploy, publish, send, and spend. It is earned separately for each lane and action class. It never attaches to a Bot as blanket trust, and authority narrows when work is delegated.

A0 — Explicit approval

The system can investigate, prepare artifacts, and assemble evidence, but a human must approve the lasting action. New lanes and new consequence classes begin here. ContextStack's own self-improvement lanes remain A0: the stack does not silently authorize changes to itself.

For a coding lane, A0 means manual review and merge. For a communications lane, it means reviewing the message before send. The policy level is the same; the action and evidence are lane-specific.

A1 — Prepared consequence, one judgment

The work and its acceptance checks complete before the operator is interrupted. The approval request contains the proposed consequence, evidence, policy, and reversal or compensation plan. The human supplies one focused judgment rather than supervising the execution.

Tap-to-merge is the coding example. Approve-to-publish and approve-to-deploy are other examples. A1 removes context assembly from the human without removing the decision.

A2 — Evidence-earned pre-authorization

A narrow action class may proceed without a contemporaneous tap only after a sustained verified record and only when its consequence is reliably reversible or compensable. Monitoring and reconciliation remain active, and one qualifying flag demotes the authorization.

Auto-merge with automatic revert is the clearest software example. A2 is not “full autonomy,” and it is not a property of the organization as a whole.

Actions without a dependable reversal, such as sending or spending, do not graduate to open-ended A2. They require explicit approval or an individually revocable pre-authorization that is bounded, idempotent, and reconciled.

Why oversight concentrates at consequence

Attention is the scarce resource. Requiring a person to watch every tool call creates fatigue without improving authority. ContextStack lets bounded work run, then concentrates judgment where state will actually change.

The human interface is delegation, exception handling, evidence review, and approval. It is not workflow programming and it is not continuous prompting.

Current status

No lane has been graduated to A2. The coding-lane simulation demonstrates the merge-specific A0/A1/A2 policy, not the complete authority model and not production autonomy. The first evidence-backed A2 lane belongs to a later horizon; containment and authority controls remain prerequisites.