Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Constrained Process Maps for Multi-Agent Generative AI Workflows

Ananya Joshi, Michael Rudow

TL;DR

The paper addresses uncertainty and coordination in regulated GenAI workflows by modeling multi-stage decisions as a bounded-horizon MDP on a directed acyclic graph, with nodes representing LLM-based reviewers and escalation paths reflecting SOPs. It introduces a multi-agent architecture where each node yields Monte Carlo samples to quantify per-agent epistemic uncertainty, while the MDP structure captures how this uncertainty propagates to system-level outcomes and human-review triggers. Empirical evaluation on the AI-safety self-harm benchmark (AEGIS 2.0) shows up to 19% accuracy gains and up to 85x reductions in required human review, with analysis of false negatives and annotation corrections highlighting the method's interpretability and tunability. The framework offers a scalable, auditable template for deploying GenAI in regulated environments, enabling targeted improvements to prompts, escalation policies, and process maps across domains.

Abstract

Large language model (LLM)-based agents are increasingly used to perform complex, multi-step workflows in regulated settings such as compliance and due diligence. However, many agentic architectures rely primarily on prompt engineering of a single agent, making it difficult to observe or compare how models handle uncertainty and coordination across interconnected decision stages and with human oversight. We introduce a multi-agent system formalized as a finite-horizon Markov Decision Process (MDP) with a directed acyclic structure. Each agent corresponds to a specific role or decision stage (e.g., content, business, or legal review in a compliance workflow), with predefined transitions representing task escalation or completion. Epistemic uncertainty is quantified at the agent level using Monte Carlo estimation, while system-level uncertainty is captured by the MDP's termination in either an automated labeled state or a human-review state. We illustrate the approach through a case study in AI safety evaluation for self-harm detection, implemented as a multi-agent compliance system. Results demonstrate improvements over a single-agent baseline, including up to a 19\% increase in accuracy, up to an 85x reduction in required human review, and, in some configurations, reduced processing time.

Constrained Process Maps for Multi-Agent Generative AI Workflows

TL;DR

The paper addresses uncertainty and coordination in regulated GenAI workflows by modeling multi-stage decisions as a bounded-horizon MDP on a directed acyclic graph, with nodes representing LLM-based reviewers and escalation paths reflecting SOPs. It introduces a multi-agent architecture where each node yields Monte Carlo samples to quantify per-agent epistemic uncertainty, while the MDP structure captures how this uncertainty propagates to system-level outcomes and human-review triggers. Empirical evaluation on the AI-safety self-harm benchmark (AEGIS 2.0) shows up to 19% accuracy gains and up to 85x reductions in required human review, with analysis of false negatives and annotation corrections highlighting the method's interpretability and tunability. The framework offers a scalable, auditable template for deploying GenAI in regulated environments, enabling targeted improvements to prompts, escalation policies, and process maps across domains.

Abstract

Large language model (LLM)-based agents are increasingly used to perform complex, multi-step workflows in regulated settings such as compliance and due diligence. However, many agentic architectures rely primarily on prompt engineering of a single agent, making it difficult to observe or compare how models handle uncertainty and coordination across interconnected decision stages and with human oversight. We introduce a multi-agent system formalized as a finite-horizon Markov Decision Process (MDP) with a directed acyclic structure. Each agent corresponds to a specific role or decision stage (e.g., content, business, or legal review in a compliance workflow), with predefined transitions representing task escalation or completion. Epistemic uncertainty is quantified at the agent level using Monte Carlo estimation, while system-level uncertainty is captured by the MDP's termination in either an automated labeled state or a human-review state. We illustrate the approach through a case study in AI safety evaluation for self-harm detection, implemented as a multi-agent compliance system. Results demonstrate improvements over a single-agent baseline, including up to a 19\% increase in accuracy, up to an 85x reduction in required human review, and, in some configurations, reduced processing time.
Paper Structure (8 sections, 2 equations, 2 figures, 2 tables)

This paper contains 8 sections, 2 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Experimental Setup
  5. Results and Discussion
  6. Conclusion
  7. Appendix
  8. Full Experimental Results and Specification

Figures (2)

  • Figure 1: Compliance frameworks that explicitly delineate agents in a process map can be promising in compliance applications. Monte Carlo simulations provide an opportunity to empirically quantify uncertainty of a labeling process with these agents that does not rely on querying LLMs about their uncertainty or setting explicit distributions on LLMs. This set-up enables transition probabilities to be learned across escalation edges and identify the agents that need to be tuned to improve output accuracy.
  • Figure 2: Complete output confusion matrices across safe and unsafe ground truth labels, and the approaches' classification (including uncertain, that's passed up for human review).