Framework

The framework begins where intent becomes action.

Agent Responsibility Engineering is a control architecture for autonomous systems in production. It governs the execution boundary: the point where an output, plan, or request becomes an act with consequences.

Control Model

ARE turns agent governance into an execution discipline.

The framework does not assume that a capable model should be allowed to act. It requires authority to be evaluated structurally before the system crosses into execution.

Intent

The agent proposes.

ARE treats intent as an input to governance, not as permission to execute.

Authority

Policy is interposed.

A governed system tests identity, scope, delegation, and context at the moment of action.

Decision

The system can refuse.

Permit, constrain, modify, or deny are engineering outcomes, not after-the-fact opinions.

Evidence

Execution leaves proof.

Every consequential action should produce a defensible record of why it was allowed or denied.

Boundary Conditions

The hard line is authority at execution.

ARE is intentionally narrower than generic responsible AI language. It asks whether a system can prove the right to perform a specific action under a specific scope at a specific time.

Capability is never treated as authority.

Approval is not enough if authority has expired before execution.

Monitoring is not governance unless the system can interpose.

Evidence is produced as part of the action path, not assembled after loss.