LedgerAgent: Structured State for Policy-Adherent Tool-Calling Agents

LedgerAgent is a newly proposed method for enhancing tool-calling agents in customer service by maintaining task states in a separate ledger, improving state management during interactions. This approach allows agents to reference current task states explicitly, thereby reducing the risk of stale or incorrect information influencing decisions and ensuring adherence to domain policies. The implementation of LedgerAgent has demonstrated improved average pass@k scores across four customer service domains, particularly excelling in scenarios requiring strict multi-trial consistency, which is crucial for practitioners developing reliable LLM applications.