Redirecting a running agent with revisioned instructions and reconciled side effects

Goal

Status as of 25 October 2026. Apply a user's correction, pause or replacement objective while preserving work whose effects are already real. Re-check the runtime's interruption semantics and the remote service's cancellation and duplicate-detection guarantees. This proposal concerns changing intent during execution; a periodic progress checkpoint alone does not establish that transition.

Prerequisites

Keep a durable task identifier, an ordered instruction revision and a ledger of attempted external actions. Each attempt needs its target, input reference, operation identifier and observed outcome. Provide an admission point that can reject actions belonging to a superseded revision. Without one, use an orchestrator that waits for affected workers to stop before releasing replacement work.

Steps

1. Classify and record the update. Distinguish a question about progress from a correction, pause, cancellation or new objective. Preserve unaffected requirements. Persist the update and its revision through a serialized control path, and acknowledge which interpretation was accepted. Ask for clarification when the intended change cannot be determined safely.
2. Close admission for affected old work. Prevent new side effects under the superseded revision before dispatching replacement work. Mark affected children and queued actions as needing reconciliation. A boundary check must be coordinated with action admission, not merely performed earlier in a worker's reasoning. Already admitted actions remain in the ledger until their outcomes are known.
3. Request interruption and observe the result. The A2A specification describes cancellation as an attempt that may fail, including when work has already completed. Record “stop requested” separately from a confirmed terminal state. Inspect tracked tasks and action receipts instead of interpreting a closed interface as proof of cancellation.
4. Resolve ambiguous effects before retrying. Classify each attempt as confirmed successful, confirmed unsuccessful or unknown. Query the authoritative remote state when available. Reuse an operation's idempotency key only for the same intended action under that service's documented contract and retention window. A2A makes duplicate detection for message sending optional; a message identifier alone is insufficient evidence of deduplication. If an unknown effect cannot be reconciled, block its retry and report the uncertainty.
5. Checkpoint the transition. Store the new objective revision, completed effects, reusable artifacts, pending operation identifiers and remaining budget. Preserve the old revision's record. Discard superseded plans from active instructions while keeping their evidence available for audit. Treat any compensating change as a separately authorized action, rather than calling it an automatic undo.
6. Resume with explicit lineage. Give each affected child the accepted revision and require that revision on its result. Under A2A's task lifecycle, a terminal task cannot restart; related follow-up work needs a new task. Reuse valid artifacts by reference, label the successor task and reject late results that do not match its scope.

Expected result

The user can see what changed, what completed before the change and what remains unresolved. Replacement work starts from reconciled state. A cancellation request never becomes evidence that an external effect disappeared, and a delayed response cannot overwrite the current objective without review.

Limits and test basis

This is an untested coordination protocol. Exercise it with a fake service that delays its response after recording an action, plus a child that returns after redirection. The proposed acceptance conditions are no blind duplicate action, no adoption of superseded instructions and an explicit unresolved state when reconciliation is impossible.