Understand Yielded Agent
Persistence and durability
Persistent history preserves a conversation. Durable execution also preserves the obligation to finish accepted work. Your host keeps one owner per Thread; Yielded Agent records progress and resumes it when that owner can run again.
Choose the guarantee
Section titled “Choose the guarantee”| Assembly | Retains history | Recovers interrupted work |
|---|---|---|
| In-memory | Until its Scope closes or the process exits | No |
| Persistent history | Successful Runs in your database | No |
| Durable host | Committed history and unfinished obligations | Yes, within the tool recovery contract |
A durable Receipt means admission is retained. The accepted submission is owed one terminal Settlement: completed, failed, or aborted. Waiting for approval or reconciling an unknown effect can leave it pending. Host availability and a usable current agent binding are still required to make progress.
Recover recorded progress
Section titled “Recover recorded progress”Host
Reopen the Thread under one owner
Storage
Load current Run state and referenced evidence
Yielded Agent
Restore input, accounting, and recorded results
Missing or invalid evidence leaves the accepted work owed with a typed failure.
Yielded Agent
Resume only work the recovery contract permits
Storage
Retain the terminal settlement
Delivery and acknowledgement may still need their own recovery.
The current Run checkpoint preserves original input and context, progress, and cumulative turn, tool, token, and cost accounting. Replacement Attempts do not receive fresh counters. Recorded results are reused. An incomplete model response may require another model call, so provider charges can repeat.
The checkpoint is mutable and deleted at settlement. It is not a historical proof of every execution transition. The journal retains committed conversation and outcomes; it provides no tamper evidence. Verification checks retained schemas, identities, ordering, references, and an active Run’s original prompt digest. It cannot attest discarded checkpoint revisions.
An uncertain booking stays uncertain
Section titled “An uncertain booking stays uncertain”Suppose a booking tool sends a reservation, then the process dies before recording the reply. The external booking may exist. Retrying blindly could book twice.
| Tool boundary | Recovery behavior |
|---|---|
| Recorded result | Reuse it |
| Ordinary mutating call with an uncertain outcome | Retain an Unknown outcome; reconcile before proceeding |
| Durable Step with a recorded result | Reuse the Step result |
| Durable Step interrupted before recording its result | May execute again; the application needs idempotency or reconciliation |
| Initial approval wait before dispatch | Preserve the wait; no handler has started |
Unfinished readonly and declared-idempotent operations can resume under compatible recorded replay semantics. An explicit Unknown outcome still requires resolution.
Exactly-once recording is not exactly-once external execution. Tool annotations must accurately describe the effect. An abort stops future work; it cannot undo an external effect or erase uncertainty. Unknown work can remain owed while later input runs in the Thread. See operations for resolution and authorization.
Keep state and deployment together
Section titled “Keep state and deployment together”Register a current binding for each stable Agent ID. Pending operations retain their original contracts; incompatible or missing bindings can leave receipts waiting for repair. Keep completion projectors pure because recovery may reevaluate them from recorded tool results.
Back up active storage with its operational state. Streamed Thread export is for idle, quiescent Threads with no unfinished Runs, approvals, unknown effects, or deliveries. Import into an empty destination. It transfers retained history and receipts, not active work. See storage and backup and restore.
An attached child owns its own Thread. The parent waits for that same child across recovery; losing an acknowledgement does not authorize starting a replacement. Background workers and messages have independent delivery lifetimes. See subagents, messaging, and the runnable durable hosts.