Calibrated Autonomy
CanonicalConfidence
Cognitive Load
Low
Evidence
production validated
Impact
feature
Ethical Guardrail
Never assume permanent high autonomy without ongoing calibration. Never surprise the user by jumping autonomy levels. Always default to the lowest safe level and escalate only when justified.
Design Intent
Autonomy is not an on/off switch -- it is a spectrum that must be calibrated to the task, the user's current trust, and the stakes involved. Calibrated Autonomy dynamically adjusts how much the agent acts on its own.
Psychology Principle
Autonomy is not an on/off switch -- it is a spectrum that must be calibrated to task, trust, and stakes.
Description
Dynamically adjust agent autonomy based on real-time signals from task risk, user trust, and confidence -- without user micromanagement.
When to use
Every agent-driven feature where the appropriate autonomy level varies by task or context.
Example
Advanced Agent Interface: Live Autonomy Dial that auto-adjusts with subtle animation + explanation: Moving to autonomous on this routine task based on your past approvals.
Autonomy Compatibility
Behavioral Objective
The agent operates at exactly the right autonomy level for the current context.
- Optimal balance of speed and control
- Reduced manual overrides over time
- Stronger long-term collaboration satisfaction
Target Actor
role
Everyday user
environment
Variable-stakes AI-assisted workflows
emotional baseline
Wants help but needs to stay in the loop
ai familiarity
medium
risk tolerance
varies
Execution Model
context_assessment
Evaluate task risk, ambiguity, and user state.
Agent uses wrong autonomy level for the situation.
dynamic_adjustment
Shift autonomy silently when safe or announce when changing.
User is surprised by sudden change in agent behavior.
user_override_and_learning
Make adjustment trivial and use it to improve future calibration.
User feels stuck at wrong autonomy level.
Failure Modes
Over-calibration creates constant micro-adjustments
Batch changes and only surface meaningful shifts
Under-calibration keeps agent too conservative
Learn from successful high-autonomy outcomes
User forgets current autonomy level
Keep the dial always subtly visible
Calibration ignores task-specific risk
Use hard ethical and risk thresholds
Learning loop is too slow
Prioritize recent user feedback
Agent Decision Protocol
Triggers
- Task context or user state changes
- Confidence or risk level shifts
- User explicitly adjusts the dial
Escalation Strategy
L1: Diagnose the failing element via behavioral_signals
L2: Nudge -- adjust copy, timing, or visual salience
L3: Restructure -- simplify flow, add progressive disclosure, restructure form
L4: Constrain -- lock Autonomy Dial to confirm_execution, add Strategic Friction
L5: Yield -- flag for human designer or domain expert review
Example
User has trusted the agent on 8 similar tasks -> agent quietly moves to autonomous mode and announces I'm handling this one fully now -- let me know if you want to review.
Behavioral KPIs
Primary
- Alignment between calibrated autonomy and user preference
- Reduction in manual overrides over time
- Collaboration satisfaction score
Risk
- Actions taken at inappropriate autonomy levels
- User frustration with autonomy mismatches
Trust
- User-reported the agent knows exactly how much to do
- Frequency of positive Autonomy Dial adjustments
Decay Monitoring
Revalidate when
- Agent model capabilities improve significantly
- User trust or risk tolerance changes
- New task types are introduced
Decay signals
- Rising manual overrides
- Drop in collaboration efficiency
- Feedback that the agent is either too hands-off or too aggressive
Pattern Relationships
Supports
Amplifies
Conflicts with