This paper introduces SagaLLM. SagaLLM is a structured multi-agent architecture designed to address four fundamental Limitations (unreliable self-verification, context loss, lack of transaction protection, and poor inter-agent coordination) of existing LLM-based planning systems. Existing frameworks leverage LLM for task decomposition and multi-agent communication, but often fail to guarantee consistency, rollback, or constraint satisfaction in distributed workflows. SagaLLM bridges these gaps by integrating the Saga transaction pattern with persistent memory, automatic compensation, and independent verification agents. It leverages LLM’s generative reasoning to automate key tasks that previously required manually coded coordination logic, such as state tracking, dependency analysis, log schema generation, and recovery orchestration. SagaLLM relaxes strict ACID guarantees, but ensures workflow-wide consistency and recovery through modular checkpointing and compensable execution. Experimental evaluations on various planning domains show that standalone LLMs frequently violate interdependent constraints or fail to recover from interruptions. In contrast, SagaLLM provides significant improvements in consistency, verification accuracy, and adaptive tuning under uncertainty, providing a strong foundation for scalable LLM-based multi-agent systems in the real world.
