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

LedgerAgent is a new inference-time method that improves how AI agents handle customer-service tasks by maintaining explicit task states in a separate ledger rather than reconstructing context from prompts. The approach reduces policy violations and improves decision consistency across multiple trials by validating state-dependent constraints before executing tool calls.

LedgerAgent addresses a fundamental architectural limitation in current tool-calling agents: the implicit management of task state. Traditional systems embed all contextual information—observations, tool returns, and policy constraints—directly into prompts, forcing agents to repeatedly extract and reconstruct relevant state from this unstructured context. This design creates predictable failure modes where agents lose track of constraints or make decisions based on outdated information.

The innovation centers on separating concerns between state maintenance and decision-making. By maintaining an explicit ledger that tracks observed facts, identifiers, constraints, and conditions, LedgerAgent creates a single source of truth that persists across conversation turns. This ledger serves dual functions: it ensures the agent's decision-making context remains accurate and fresh, and it acts as a gatekeeper that validates whether proposed tool calls comply with state-dependent policies before execution.

Across four customer-service domains and various model architectures (both open and closed-weight), LedgerAgent demonstrates measurable improvements in pass@k metrics, with particularly strong gains under stricter consistency measurements that require correct behavior across multiple trials. This suggests the approach reduces not just average performance variance but systematically prevents recurring mistakes.

For the broader AI agent ecosystem, this work highlights how explicit state management can improve reliability without requiring model retraining or fundamental architectural changes. As enterprises deploy agents in regulated domains where policy adherence is non-negotiable, techniques like LedgerAgent become critical infrastructure. The method's compatibility with both open and closed models indicates broad applicability.

• →LedgerAgent maintains explicit task state in a separate ledger to replace implicit state reconstruction from prompts
• →The approach validates state-dependent policy constraints before tool execution, preventing policy violations
• →Improvements in pass@k metrics are strongest under multi-trial consistency requirements, indicating reduced recurring errors
• →The method works across both open and closed-weight models without requiring retraining
• →Explicit state management represents a critical pattern for deploying agents in regulated customer-service domains