Glossary

Key ideas#

• Prefer stable contracts and machine-actionable statuses.
• Idempotency and retries are inseparable in distributed systems.
• Observability fields are part of the interface, not optional extras.

Diagram#

Vocabulary map

Agent -> decides goals, delegates execution
Orchestrator -> routes and coordinates workers
Worker -> bounded execution with a contract
Validator -> checks outputs before side effects
Aggregator -> merges partial outputs
Artifacts -> stored outputs referenced by ID/URL

Terms#

• AI Worker: A bounded execution unit with explicit inputs/outputs, constraints, and observability.
• AI Agent: An autonomous loop that plans and chooses actions toward a goal.
• Orchestrator: A coordinator that routes requests, schedules workers, and handles retries/fallbacks.
• Contract: The agreed interface: schemas, status codes, and semantics for failures.
• Constraint: A hard limit such as timeout, tool allow-list, rate limit, or budget cap.
• Idempotency: The property that repeated processing of the same request does not create duplicate side effects.
• Retry: A re-attempt policy, typically with backoff, for transient failures.
• Backoff: A delay strategy for retries (often exponential with jitter) to reduce thundering herds.
• Artifact: A file or payload produced during execution, stored out-of-band and referenced from the response.
• Trace: Correlated telemetry across steps; includes a trace_id and spans per worker.
• Sandbox: A restricted execution environment (filesystem/network/tool permissions) to reduce blast radius.
• DAG: A directed acyclic graph representing a pipeline of steps with dependencies.

See also#

• what is AI worker
• AI worker architecture
• AI worker design patterns
• examples