Multi-Agent Legacy Architecture

MALA Framework: A Framework for Multi-Agent AI Integration in Legacy Industrial Systems

This work introduces MALA (Multi Agent Legacy Architecture), a supervised multi-agent framework designed for industrial environments operating on heterogeneous legacy systems that lack modern APIs.

Research Overview

Many industrial environments continue to rely on terminal-based enterprise systems, flat-file archives, scanned documents, and local operational logs accumulated over decades. These systems often expose limited interoperability and present barriers to contemporary AI integration.

MALA proposes a three-tier architecture for grounded AI-assisted decision support in such settings. Rather than replacing legacy software, the framework treats existing industrial systems as operational surfaces and integrates:

• Gatherer Agents implementing Recursive Context Enrichment (RCE) to collect and normalize evidence across heterogeneous sources.
• Critic Agents enforcing deterministic safety constraints through a Constrained Markov Decision Process (CMDP) formulation.
• Human-on-the-Loop (HotL) oversight for escalation, exception handling, and verification of medium-confidence or high-risk actions.

The framework incorporates freshness-aware retrieval, evidence traceability, confidence-based escalation, ontological mapping of local industrial terminology, and auditable workflow logging for supervised deployment in industrial contexts.

Evaluation Summary

The accompanying study evaluates MALA using a retrospective replay and shadow-execution protocol on archived disruption-management cycles from a legacy European steel-production environment. The evaluation focuses on supervised decision support under matched evidence conditions and explicitly excludes autonomous live deployment.

Across a 50-cycle replay dataset, the framework demonstrated:

• Reduced decision-support latency relative to manual workflows.
• Zero recorded safety violations under CMDP-constrained execution.
• Limited Human-on-the-Loop intervention requirements.
• Improved task success relative to unconstrained fully agentic baselines.

The study additionally includes ablation analysis, threshold sensitivity experiments, grounded case-trace evaluation, and deployment-oriented discussion for industrial legacy modernization contexts.

Repository Contents

This archive may include:

• Research manuscript and supplementary material.
• Reference implementation of the MALA framework.
• Example orchestration and safety-control modules.
• Experimental configurations and replay methodology.
• Documentation for reproducibility and future extensions.

Important Notes

• Research status: Preprint / independent research artifact.
• Evaluation mode: Retrospective replay and shadow execution. No autonomous live plant actuation was performed.
• Industrial data: Any operational data included in this archive has been anonymized or abstracted where required.
• Intended audience: Researchers and practitioners in industrial AI, multi-agent systems, legacy modernization, human-AI collaboration, and safe autonomous systems.

Keywords

Multi-Agent Systems; Industrial AI; Legacy Systems; Human-on-the-Loop; CMDP Safety; Recursive Context Enrichment; AI for Manufacturing; Legacy Modernization; Explainable AI; Industrial Decision Support