durable-agent-workflowslisted

# Durable Agent Workflows Build workflows that can resume without silently repeating externally visible work. Treat orchestration durability and side-effect safety as separate properties: a checkpoint can make a workflow resumable while still allowing an activity or tool to run more than once. ## When to Use - An agent or job is retried after timeouts, worker loss, deploys, or checkpoint recovery. - A tool can charge, send, publish, provision, trade, mutate records, or start costly work. - A long-running tool may finish after its caller has timed out. - A workflow needs human approval, durable waits, or crash recovery. - Duplicate dispatches or uncertain remote outcomes must be diagnosed and tested. ## When Not to Use - A short, read-only operation where duplicate execution is harmless and inexpensive. - A request to retry an unknown production mutation immediately; reconcile it first. - A request for exactly-once delivery as a transport guarantee. Design for repeated delivery and idempotent effects instead. ## Prerequisites Before implementation, identify: - the durable runtime or checkpoint store and its retry/timeout behavior; - the authoritative database for operation records; - provider support for idempotency keys or status lookup; - business identifiers that represent one logical intent; - retention, audit, and manual-reconciliation owners. Never place credentials, authorization headers, payment data, or full sensitive tool arguments in operation keys or logs.