Agent Autonomy Level Mapper

Autonomy is not a product-wide setting. It is a decision made per action, and it should be earned.

What this skill is

A decision workflow that breaks an agent's job into discrete actions and assigns each one an autonomy level from a five-level ladder. Each action is scored on stakes, reversibility, and the agent's measured confidence, and the score maps to a level with a matching approval-gate design. The output is an autonomy map product, design, engineering, and risk teams can agree on and revisit as the agent proves itself.

What it solves

• One global "autonomous mode" toggle that is too risky for some actions and too timid for others
• Approval prompts on trivial actions that train users to click "Approve" without reading
• Irreversible actions (sending, paying, deleting) running without a gate
• Autonomy decisions made on demo confidence rather than measured accuracy
• No defined path for an agent to earn more autonomy, or to lose it after incidents

When to invoke

• Designing a new agent that can call tools, write data, or contact people
• Stakeholders disagree on whether an action needs human approval
• Users complain about too many confirmations, or about surprises
• Expanding an assistant from suggesting to acting
• Post-incident review of an agent action that should have been gated

Phase 1: Decompose the job into actions

List every discrete action the agent could take. Split compound actions: "Process refund" becomes look up order, calculate amount, draft customer message, issue refund, send message.

Read-only actions (search, summarize, look up) still go on the list; they are usually the first candidates for silent autonomy.

Phase 2: Score each action

Score 1 (low) to 5 (high) on each axis:

Confidence must come from evaluation data on this action type, not from the model's self-reported certainty. Unmeasured equals 5.

Autonomy risk = S x R + 2 x C
Range: 3 to 35

Multiplying stakes by irreversibility reflects that a high-stakes action with a reliable undo is far safer than one without.

Phase 3: Map risk to autonomy level

Override rules, applied after scoring:

• Any action with R = 5 (cannot be undone) is never placed above L3, regardless of confidence
• Actions affecting people outside the organization are capped at L4
• Regulatory or contractual requirements override the score in either direction
• If a user explicitly asks for an action in the moment, the agent may move up one level for that instance only

Phase 4: Specify the approval gates

For every L3 action, and every L4 undo window, define the gate:

Undo windows for L4: 5-30 seconds for messages, up to 24 hours for reversible data changes. Name the window length in the notification.

Phase 5: Define promotion and demotion criteria

Autonomy is earned with evidence. For each action, set the rule to move one level:

Output

• Action inventory with scope and affected parties
• Scoring table (S, R, C, risk score) per action
• Autonomy map: each action assigned L1-L5 with override notes
• Approval-gate specification for every L3 and L4 action
• Promotion and demotion criteria with the metrics to instrument
• Open questions for risk, legal, or policy review

Operating rules

Always

• Score actions individually, never the agent as a whole
• Use measured accuracy for the confidence axis; treat unmeasured as lowest confidence
• Show exact values and recipients in every approval gate
• Default to doing nothing when an approval times out
• Let users reduce autonomy for themselves without friction

Never

• Let model self-reported confidence substitute for evaluation data
• Place an irreversible external action at L5
• Gate trivial actions just to appear safe; it creates approval fatigue
• Promote autonomy without a defined metric and threshold
• Ship act-and-report without a working undo