From Sagas to SagaLLM: Long-Lived Transactions and the Agent That Writes Them

The transaction concept was built for short, flat operations written by people. Long-lived activities broke it almost at once. The database community spent two decades answering. Sagas added compensation. ConTracts added invariants and a durable context. Workflow systems added the script that ties steps together, and the transaction-processing monitor made all of it run. Each answer assumed something the others also assumed. The steps were known in advance. Their meaning was fixed. The authority to run them was settled before the first step began. Language-model agents remove those assumptions at once. The agent now writes the operation, its justification, and often its compensation. This paper traces the lineage from Gray's 1981 virtues and limitations through Sagas, ConTracts, and workflow. It arrives at the author's SagaLLM, ALAS, Mnemosyne, TRACE, and TRW, and at the collaboration contract of the third volume of his trilogy. It names the five conditions every classical guarantee took as given and an agent does not supply. It then shows what each modern component inherits from the lineage and what it adds. A durable workflow can execute the wrong authority perfectly. The old machinery is still the right machinery. What is new is the gate in front of it.