Orchestration
The state transitions, scheduling, branching, waiting, retry, cancellation and compensation that coordinate multi-step work.
Updated 15 September 2026
In practice
A workflow checkpoints after each completed step, pauses for a human approval, resumes when it arrives, and routes a timed-out payment to reconciliation instead of retrying it.
The distinction
Orchestration records what the run believes and what it should do next. It cannot prove what an external provider received, committed or rejected during a transport failure — that is a separate kind of durability.
The double-durability seam
The most important seam in agent execution runs between two different kinds of durability.
| Orchestration durability | Effect durability |
|---|---|
| Run identity | Logical effect identity |
| Checkpoint and event history | Intent record |
| Lease and worker ownership | Idempotency key |
| Pause and resume | Provider receipt or status |
| Retry and cancellation | Reconciliation and compensation |
Suppose a worker sends an update and loses its connection before the response. A checkpoint taken before the call proves the step was ready. It does not prove the update happened. A checkpoint after the call never got written.
The safe path records an intent with a stable key, queries the destination for status, and holds the run in reconciliation until the effect is known. A retry without that seam duplicates the consequence.
Why "exactly once" needs qualification
A runtime may execute a workflow step once in its own history while the external request is delivered more than once, unless the destination or the application supplies idempotency.
Durable infrastructure reduces uncertainty. It does not remove the need for a domain-level record of what the outside world was asked to do.
Sources & context
LangGraph persistence
LangChain · 2026
Documents checkpointing graph state by thread, with interrupts, resume and retry.
Source checked 2026-09-13
Microsoft Agent Framework
Microsoft · 2026
Documents workflow checkpoints capturing executor state and pending requests, with a durable extension adding persistent sessions and human-in-loop waits.
Source checked 2026-09-13