Read the Manifesto
Explore the complete principles and rationale behind the discipline.
Open the Manifesto →ARE overview
What Is Agent Reliability Engineering?
The discipline for engineering autonomous AI agents whose behavior remains measurable, bounded, observable, recoverable, and accountable in production.
Definition
Agent Reliability Engineering (ARE) is the software engineering discipline for making autonomous AI agents reliable, governable, observable, recoverable, and accountable in production.
ARE treats reliability as an architectural property—designed into an agent's identity, authority, contracts, controls, and recovery paths before it receives production authority. The objective is not to eliminate every possible failure. It is to make agent behavior measurable, bounded, explainable, containable, and recoverable when failures occur.
An agent is not reliable merely because it completes a task successfully. It must also produce acceptable outcomes through an authorized process, remain inside policy and resource boundaries, leave enough evidence to reconstruct consequential actions, and fail without creating uncontrolled impact.
Why autonomous AI agents need reliability engineering
Traditional software usually executes deterministic or bounded instructions. Autonomous agents introduce variable behavior while selecting tools, planning actions, accessing data, delegating work, and changing external systems.
Probabilistic behavior
The same task can produce different plans, tool calls, costs, and outcomes across repeated runs.
Consequential action
Agents can send messages, change records, execute code, spend money, or modify infrastructure—not only generate text.
Dynamic environments
Models, prompts, tools, data, dependencies, policies, and user behavior change after deployment.
Delegated authority
An agent may inherit permissions from users, services, tools, or other agents, creating authority that is difficult to see or control.
Compound failure
Retries, loops, handoffs, and multi-agent workflows can turn a small error into repeated or irreversible impact.
Accountability requirements
Operators need to know which identity acted, what authority applied, why the action was allowed, and how it can be corrected.
What reliable agent behavior requires
Acceptable outcomes
The agent completes its intended task with sufficient accuracy, consistency, and quality.
Enforced boundaries
Actions, data access, budgets, delegation, and prohibited operations remain inside explicit authority and policy.
Operational evidence
Consequential actions retain identity, authority, inputs, decisions, tool use, policy verdicts, outputs, and outcomes.
Bounded failure and recovery
Unsafe behavior can be stopped, contained, reversed, compensated for, escalated, or resumed safely.
How Agent Reliability Engineering works
| ARE concept | Engineering question | Purpose |
|---|---|---|
| AI Agent Reliability | Does the agent behave acceptably and recover safely? | Defines the measurable outcome ARE is trying to produce. |
| Reliability Surface | Which behavioral properties and controls determine trustworthiness? | Maps consistency, robustness, predictability, safety, identity, authority, contracts, policy, testing, runtime controls, observability, governance, and improvement. |
| Agent Authority | What may the agent do, to which data, under whose permission? | Separates technical capability from explicit and enforceable permission. |
| Agent Governance | Who owns the risk, policies, evidence, and operating decisions? | Connects technical enforcement to organizational accountability. |
| Agent Reliability Controls | How are unacceptable actions prevented, detected, contained, and recovered? | Turns reliability requirements into executable boundaries. |
| Reliability Debt | Is capability or authority growing faster than reliability engineering? | Makes accumulating operational risk visible and actionable. |
| Progressive Autonomy | What authority is justified by current evidence? | Expands autonomy as reliability improves and contracts it when reliability deteriorates. |
ARE and related disciplines
| Discipline | Primary focus | Relationship to ARE |
|---|---|---|
| Site Reliability Engineering | Availability and operation of software services and infrastructure. | ARE adapts reliability thinking to agents whose behavior, plans, and tool use are probabilistic. |
| AgentOps | Deployment, monitoring, evaluation, and operation of agent systems. | ARE adds an explicit model for authority, governance, failure containment, recovery, and reliability evidence. |
| AI safety | Preventing harmful outcomes from AI systems. | ARE implements production controls that help keep agent behavior inside defined safety boundaries. |
| AI security | Protecting models, data, tools, identities, and infrastructure. | Security controls form part of the Reliability Surface but do not by themselves establish operational reliability. |
| AI governance | Policy, accountability, risk, compliance, and oversight. | ARE translates governance requirements into enforceable runtime controls and evidence. |
| AI evaluation | Measuring model and agent behavior. | Evaluation supplies evidence, while ARE connects that evidence to controls, recovery, and operating authority. |
ARE does not replace these disciplines. It connects their relevant engineering practices around autonomous systems operating in production.
A practical example
Consider an AI support agent that can draft responses, issue refunds, change customer records, and contact users. Task accuracy alone does not make the agent reliable. The system must also establish which customers and records it may access, how much it may refund, which actions require approval, how every decision is recorded, and how an incorrect action can be reversed.
Under ARE, the agent begins with restricted authority. Its proposed actions are observed and evaluated before execution. It may then advance to drafting actions, acting with explicit approval, and eventually operating autonomously inside a narrow, verified scope. If policy violations, recovery failures, drift, or customer impact cross defined thresholds, its authority contracts to a safer operating stage.
The goal is not maximum autonomy. The goal is justified autonomy: authority that remains proportional to measured reliability and can be revoked without redesigning the system.
Applying ARE to an agent system
- Inventory the agent's identity, tools, data, credentials, budgets, delegation paths, and downstream effects.
- Define acceptable outcomes, explicit authority, prohibited actions, contracts, and policy boundaries.
- Evaluate realistic tasks, edge cases, policy compliance, known failure modes, and recovery behavior.
- Implement preventive, detective, containment, observability, and recovery controls outside the prompt where possible.
- Grant authority progressively, monitor reliability evidence, and contract authority when conditions deteriorate.
ARE is relevant to software, platform, security, governance, risk, operations, and product teams responsible for autonomous systems. It is an open working draft, not a formal industry standard, and is developed through public review and contribution.
Continue exploring Agent Reliability Engineering
Explore the concepts
Review the canonical definitions for Reliability Surface, Reliability Debt, Progressive Autonomy, authority, governance, and controls.
View all concepts →Assess reliability maturity
Use the five-level model to connect engineering evidence to operating authority.
Open the maturity model →Review the principles
See the architectural and operational commitments that guide ARE implementation.
Read the principles →