FRAMES: A New Framework for Evolving LLM Agents in Enterprise Workflows
A new framework named FRAMES has been developed by researchers to enhance LLM agents functioning within policy-driven enterprise processes, including document auditing. This framework tackles significant issues such as limited and unlabelled operational feedback, potential regressions when rules are modified, and the necessity to boost accuracy without raising inference costs or sacrificing auditability. FRAMES initiates deployable skills from existing resources and refines them through consensus-based mutation, Pareto selection balancing accuracy and cost, and an anti-regression assurance, all while maintaining auditability. When implemented in an internal production environment, FRAMES demonstrated the optimal accuracy-cost balance compared to other baselines, with similar improvements confirmed on tau-bench. The research is documented on arXiv with the identifier 2608.01772 and is categorized under Computer Science and Artificial Intelligence.
Key facts
- FRAMES is a closed-loop framework for evolving LLM agents in policy-governed enterprise workflows.
- It addresses sparse and unlabeled operational feedback in document auditing.
- It uses consensus-based mutation and Pareto selection over accuracy and cost.
- It includes an anti-regression guarantee to prevent rule edits from causing regressions.
- It preserves auditability throughout the evolution process.
- FRAMES was deployed on an internal production system, achieving the best accuracy-cost trade-off among baselines.
- The same gains were reproduced on tau-bench.
- The paper is available on arXiv with identifier 2608.01772.
Entities
Institutions
- arXiv