Applied Autonomy Framework

Autonomy Borders

Every AI agent operates within boundaries. An autonomy border is the point at which an agent's independence ends — where it must stop, escalate, or hand off. The border architecture determines how autonomy is bounded in practice.

Core Concept
Three Layers of Autonomy Borders
Designed Borders Set at deployment. Fixed rules.
Conditional Borders Shift at runtime. Input-dependent.
Learned Borders Develop over time. Feedback-driven.
Agent
Border Types

Designed Borders

Set at build / deploy time

Deliberate governance decisions. Do not change during execution. Represent what the agent should never do on its own.

Agent restricted to approved toolset. Cannot access financial systems. Must route legal questions to human.

Conditional Borders

Shift at runtime per inquiry

Same agent, same profile — but effective autonomy changes based on properties of the specific work being handled.

Expense agent: autonomous under $5K, human approval above $5K. Same agent, different border based on amount.

Learned Borders

Develop through feedback over time

Agent compares current work against accumulated human annotations and feedback. Calibrated confidence earned through guidance.

Agent matches situation to patterns where humans previously corrected or redirected it. Judgment improves with experience.
Border Response
What Happens at the Border?
When an agent recognises it has reached an autonomy border, three responses are possible.
Trigger
Agent reaches autonomy border

Stop

Halt and communicate why. Cost of wrong action exceeds cost of inaction.
Safe default

Escalate

Transfer decision with context to a human or another agent. Preserves progress.
Preferred

Push Through

Proceed despite the border. Usually a failure mode — confidence exceeds competence.
Governance risk
Across Dimensions
Every Dimension Has Its Borders
Autonomy borders manifest differently in each of the four autonomy dimensions.

Task

"Is this request within my domain?"
Border hit: Customer service agent receives a legal question. Recognises this is outside its scope.

Tool

"Do I have the right capabilities?"
Border hit: Agent needs database access it doesn't have. Available tools are insufficient for the task.

Plan

"Can I find a viable path forward?"
Border hit: Agent's approach has failed repeatedly. Problem is more complex than its strategies can handle.

Collaboration

"Do I need help, and from whom?"
Meta-border: Border events in other dimensions often resolve here — through escalation.

Collaboration as Meta-Border

Most autonomy border events ultimately resolve through collaboration. An agent that hits a task, tool, or plan border typically responds by crossing a collaboration border — escalating to a human or delegating to another agent. This makes collaboration the connective tissue of the border architecture.