FRAMES: Guarded and Dual-Objective Skill Evolution for Agents in Policy-Governed Enterprise Workflows

As LLM agents are increasingly deployed to handle complex, policy-bound enterprise workflows—such as financial document auditing—they face significant challenges: they must maintain high accuracy, adhere to strict regulatory policies, and remain auditable, all while keeping inference costs manageable. The authors investigate how to systematically improve these agents' performance without retraining the underlying models, specifically addressing the difficulty of sparse feedback, the risk of regression when updating rules, and the need for cost-efficiency.

FRAMES (Feedback-driven Reasoning with Adaptive Mutation and Evolution of Skills) is a closed-loop framework that operates in two phases. First, it performs a cold start by converting enterprise policies and available operational feedback into a structured, deployable skill bank. Second, it employs an iterative evolution loop that uses a population of independent diagnosers to identify root causes of failures. These diagnoses are consolidated into skill edits, which are then subjected to a per-category non-regression gate and a Pareto-based selection process to balance accuracy and inference cost. This ensures that improvements are both effective and compliant with existing business rules.

When deployed on a production financial document-auditing system (FinDAS), FRAMES demonstrated superior performance compared to several baselines, including manual human editing and existing automated skill-evolution methods. It achieved the best accuracy-cost trade-off, showing significant improvements in handling hallucination and edge cases. The results were further validated on the public tau-bench, confirming that the framework's gains generalize beyond the internal production environment. The system's ability to maintain a versioned, auditable trail of natural-language edits makes it particularly suitable for regulated enterprise settings.