What is Workflow Orchestration?
Workflow Orchestration is the structured coordination of tasks, dependencies, state, timers, signals, policies, approvals, external effects, and terminal outcomes across a bounded business process that may include models or agents.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Workflow Orchestration treats model and agent calls as activities inside a larger business process. The workflow owns the process identity, input and output contracts, state transitions, dependencies, schedules, approvals, retry classes, cancellation, and terminal status. An LLM may classify, extract, generate, or choose a bounded route, but it does not replace the orchestrator or the source system's authorization.
A chain is an ordered composition of steps. A graph expresses explicit branches, cycles, joins, and state. Workflow Orchestration is the operational discipline that runs those structures across worker failure, long waits, deploys, and external events. An Agentic Workflow is a workflow that delegates selected runtime choices to agents; deterministic workflows remain valuable for repeatable or regulated processes.
Durable execution usually separates replayable orchestration from non-deterministic activities such as model calls, network requests, clocks, random values, and writes. A runtime may rebuild state from an event history or resume from checkpoints, so workflow code and version changes must preserve replay compatibility. Completed activity results may be reused during replay, but activity dispatch can still be repeated around failure boundaries. External writes therefore require stable idempotency keys, effect status, and downstream reconciliation.
Retry policy belongs to each failure class, not the workflow globally. Retry transient provider or network failures with backoff, jitter, and deadlines. Route validation and policy failures to repair, rejection, or review. A timeout after dispatch can be outcome_unknown; do not label it failed or blindly repeat it. If prior committed actions must be undone, use explicit compensation or a Saga. Compensation is a new fallible action, not time travel or a database rollback.
Human approval and external callbacks should be durable signals correlated to one workflow run and exact proposed operation. Persist actor, approver, resource, argument and schema digests, policy version, decision, and expiry; verify them again before execution. Waiting should not depend on a live web request or worker process.
Version workflow definitions, state schemas, activities, prompts, models, policies, and payload contracts. Test old histories or checkpoints against new code, define migration or drain behavior for in-flight runs, and bound event-history and state growth. Observability should expose workflow and run IDs, step attempts, decisions, waits, effects, costs, and terminal reasons without recording hidden chain-of-thought or unnecessary private data.
Evaluate both control-flow correctness and business outcome. Cover duplicate events, out-of-order signals, stale approval, worker crash, provider timeout, cancellation, concurrent update, schema migration, replay divergence, compensation failure, and poison payloads. Durable execution improves recovery mechanics; it does not prove that a model decision, tool result, or business action is correct.
Key Characteristics
- Bounded process identity, contracts, dependencies, state, signals, timers, and terminal outcomes
- Replayable orchestration separated from non-deterministic model, tool, network, and write activities
- Per-failure retry, timeout, cancellation, idempotency, unknown-outcome, and compensation policies
- Durable human approval and callbacks correlated to exact workflow operations
- Versioned definitions, state schemas, activities, prompts, models, policies, and migrations
- Operational evidence for attempts, waits, decisions, effects, costs, recovery, and final outcome
Common Use Cases
- Document intake, extraction, validation, human review, indexing, and publishing
- Scheduled prompt, model, retrieval, and safety evaluation pipelines
- Long-running agent tasks that survive worker restarts and approval waits
- Business processes with idempotent writes and explicit compensation steps
- Batch AI jobs with concurrency limits, fan-out, joins, retries, and traceable outputs
- Regulated processes where model decisions are bounded by deterministic policy gates
Example
Loading code...Frequently Asked Questions
How does Workflow Orchestration differ from Agentic Workflow?
Workflow Orchestration is the broader process discipline and may be fully deterministic. An Agentic Workflow delegates selected runtime decisions to an agent while the orchestrator still enforces state, policy, budgets, effects, and terminal outcomes.
Does durable execution guarantee exactly-once side effects?
No. A runtime can replay orchestration and persist completed activity results, but an activity may be dispatched again near a failure boundary. External writes still need idempotency keys, effect records, and downstream reconciliation.
What is the difference between retry and compensation?
Retry attempts the same operation after a classified transient failure. Compensation performs a separate business action to counter an earlier committed effect. Compensation can fail and must itself be observable and idempotent.
How should workflow upgrades handle in-flight executions?
Replay old histories against new code before release. Preserve deterministic behavior through version markers, compatible activities, state migration, worker version routing, or draining old runs. Do not assume new code can replay every old state.
Does Workflow Orchestration make AI output reliable?
It improves execution reliability, not semantic correctness. Model outputs still need schema validation, evidence checks, policy gates, adversarial evaluation, human review where required, and verification against authoritative systems.