Read vs write
Separate exploratory reasoning from actions that change external systems.
Agent Reliability
Design AI agents to fail safely
Agent failures become operational incidents when actions are irreversible, retries duplicate side effects, or a model continues despite missing evidence. Recovery must be part of the workflow design.
Answer in brief
Bound every action, make tools idempotent where possible, checkpoint durable state, use circuit breakers and approval gates, and test rollback and resume behavior under partial failure.
Failure boundaries
Use these dimensions to make the architecture decision explicit and testable.
Blast radius
Limit scope, spend, records, and time per execution.
Permissions
Issue short-lived, task-specific credentials.
Safe state
Define the condition the system returns to after interruption.
Recovery mechanics
Use these dimensions to make the architecture decision explicit and testable.
Checkpoint
Persist plan, state, approvals, and completed side effects.
Idempotency
Use request keys and deduplication so retries do not repeat writes.
Compensation
Define reversible actions or compensating transactions.
Circuit breaker
Stop repeated failure and require review before resuming.
Verification
Use these dimensions to make the architecture decision explicit and testable.
Preconditions
Validate inputs, permissions, and external state immediately before action.
Postconditions
Confirm the intended change actually occurred.
Replay tests
Inject timeouts, malformed responses, stale state, and tool outages.
Operations
Expose pause, resume, cancel, inspect, and rollback controls.
Production checklist
Use this as a design-review gate before release.
- Classify side effects by reversibility.
- Set execution and spend limits.
- Use idempotency keys.
- Checkpoint before consequential actions.
- Test compensation and rollback.
- Require human approval when recovery is uncertain.
Continue learning
Place this pattern inside a broader agent architecture and operating model.
AI agent tutorial library
Browse memory, planning, frameworks, reliability, safety, and observability.
Open the tutorial library →AI agent guide
Start with agent features, design patterns, benefits, challenges, and industry applications.
Open the guide →