The Autonomy Dial
CanonicalConfidence
Cognitive Load
Low
Evidence
production validated
Impact
platform
Live Preview
Agent drafts the response. You review and decide whether to send.
Ethical Guardrail
User must always be able to dial back to suggest_only with one click. Certain safety-critical actions lock the dial at confirm_execution regardless of user preference.
Design Intent
Human agency is non-negotiable in product workflows where accountability attaches to decisions. The Autonomy Dial gives users a single, always-visible control that governs how much latitude the AI agent has. At its core, it answers the question every user silently asks: 'How much do I trust this thing?' By making trust a continuous, adjustable setting rather than a binary on/off, users can ratchet up autonomy as calibrated trust builds -- and instantly pull back when the stakes change. For the agent, the dial is law: it defines the ceiling of what the agent may do without explicit human approval.
Psychology Principle
Users need graduated control over AI agency to maintain psychological ownership of outcomes.
Description
A visible interface control that lets the user choose how much the AI agent may act on their behalf. Three positions: Suggest Only (agent drafts, user decides), Confirm Execution (agent proposes a plan, user approves), and Autonomous (agent acts, user reviews after). This is the central mechanism for maintaining human agency in an agent-first product.
When to use
Every agent-driven feature. The dial should be present (or its equivalent logic should govern) any interaction where the agent could take action.
Example
GitHub/Linear: 'Draft issue description for me' (Suggest Only) vs. 'Triage and label this issue' (Confirm Execution) vs. 'Auto-assign and prioritize incoming issues' (Autonomous).
Autonomy Compatibility
Behavioral Objective
Users maintain felt ownership of outcomes even as the agent takes on more work.
- Users consciously choose their autonomy level rather than accepting a default
- Users increase autonomy gradually as trust is earned through correct agent actions
- Users can instantly reduce autonomy when entering high-stakes or unfamiliar workflows
Target Actor
role
Any product user interacting with agent features
environment
Product development workflows ranging from routine (daily standups, status updates) to high-stakes (production deployments, payment authorizations)
emotional baseline
Ranges from cautious first-time users to confident power users who want maximum throughput
ai familiarity
low-to-high (dial must serve the full spectrum)
risk tolerance
variable by task context
Execution Model
present
The autonomy dial is rendered as a visible, always-accessible control in the agent interface. Three discrete positions are clearly labeled.
User cannot locate the dial or does not understand what the positions mean.
calibrate
System sets a sensible default level based on workflow context, user history, and risk profile. High-stakes workflows default to Confirm Execution. First-time users default to Suggest Only.
Default is inappropriate for context (e.g., Autonomous on a first-time security review).
operate
Agent acts strictly within the permissions granted by the current dial position. Suggest Only means the agent drafts but never submits. Confirm Execution means the agent prepares and presents a plan for approval. Autonomous means the agent executes and reports after.
Agent takes action beyond the permitted level.
override
User changes the dial level at any time. Change takes effect immediately. If moving to a higher autonomy level on a high-stakes workflow, a brief confirmation is shown.
User cannot change the dial, or change does not take effect immediately.
lock
System locks the dial at a specific level for safety-critical workflows. User is informed why the lock exists and cannot override it. Lock is logged in the Audit Trail.
User bypasses a safety lock, or lock is applied without explanation.
Failure Modes
Users leave the dial at default and never consciously choose
Prompt dial selection on first use of each workflow type. Show brief onboarding tooltip.
Agent exceeds permitted autonomy level due to race condition
Server-side permission enforcement. Dial state is authoritative, not advisory.
Power users resent safety locks
Always explain why the lock exists with specific policy reference. Frame as protection, not restriction.
Users oscillate rapidly between levels, creating unpredictable agent behavior
Debounce dial changes. Queue pending agent actions during transition.
Agent Decision Protocol
Triggers
- New workflow entered (calibrate default)
- User changes dial level (adjust agent behavior immediately)
- Safety rule matched (lock dial)
- Trust erosion signals detected (suggest dial reduction)
Escalation Strategy
L1: Respect -- agent operates strictly within dial-permitted level
L2: Nudge -- if trust erosion signals detected, suggest reducing autonomy level
L3: Restructure -- on repeated overrides at Autonomous, prompt user to consider Confirm Execution
L4: Constrain -- safety lock engages, dial fixed at confirm_execution
L5: Yield -- flag workflow for supervisor review if dial conflicts with risk level
Example
New user starts first code review workflow -> dial defaults to Suggest Only -> agent drafts review comments, user approves -> after 5 successful reviews, user moves to Confirm Execution -> agent now proposes full review plan, user approves with one click -> security-sensitive repo detected -> dial locks to Confirm Execution with explanation.
Behavioral KPIs
Primary
- % of users who actively set dial level (vs. accepting default)
- Average autonomy level over time per user (trust trajectory)
- Time from onboarding to first Confirm Execution usage
Risk
- Agent actions exceeding permitted level (should be 0%)
- Safety lock bypass attempts
Trust
- Dial regression frequency (moving from higher to lower autonomy)
- Correlation between dial level and task completion rate
Behavioral Signals
trust_erosion
dial_regression_count > 2 in 7 days
suggestion_dismissal_rate > 60%
autonomy_mismatch
user_override_at_autonomous_level=true (user reverses agent action taken at Autonomous)
dial_level=autonomous AND revert_count > 0 in session
Decay Monitoring
Revalidate when
- New autonomy level requested (e.g., 'Semi-Autonomous' between Confirm and Full)
- Security policies change requiring new lock rules
- User research reveals confusion about dial positions
Decay signals
- Declining active dial selection rate
- Increasing trust erosion signals across user base
- Rising support tickets about agent doing too much or too little
Pattern Relationships
Amplifies
Conflicts with