ARE principles
Principles of Agent Reliability Engineering
Reliability is an architectural property. These principles translate that claim into decisions teams can make before and during production operation.
1. Governance is architectural, not operational
Identity, authority, policy enforcement, evidence, and accountability must exist before the first consequential action. Governance built only as a meeting or review process will be bypassed when it becomes inconvenient.
2. Authority is explicit, verifiable, and earned
An agent's ability to call a tool is not permission to use it unsupervised. Every action should be traceable to an enforceable authority scope. Expanded authority requires evidence.
3. Reliability must be observable
A team cannot manage reliability it cannot see. Consequential actions require enough context to reconstruct identity, authority, inputs, tool use, policy decisions, outputs, and recovery.
4. Failure must be bounded and recoverable
Probabilistic systems will fail. Runtime controls should limit duration, cost, repetition, blast radius, and irreversible effects while supporting retry, rollback, compensation, or controlled termination.
5. Autonomy is reversible
Authority expands as reliability improves and contracts when it deteriorates. Systems should be designed to return to a safer operating stage without redesigning the entire workflow.
6. Critical policy is enforced outside the agent
Prompts can express intent, but they are not the sole enforcement boundary for permissions, spending limits, data access, approvals, or prohibited actions.
7. Improvement is continuous and controlled
Model, prompt, tool, policy, and environment changes can alter behavior. Evaluation, testing, incident learning, and configuration versioning should make improvement measurable and reversible.