[Paper Review] BMW Agents -- A Framework For Task Automation Through Multi-Agent Collaboration

N Crawford, Edward Duffy+6 more|arXiv (Cornell University)|Jun 28, 2024

Multi-Agent Systems and Negotiation4 citations

TL;DR

BMW Agents presents a scalable, modular multi-agent framework for industrial task automation using Large Language Models (LLMs), enabling collaborative workflows through structured planning, tool integration, and episodic memory. The framework supports complex, real-world applications via reusable agent units, human-in-the-loop feedback, and restartable execution, achieving robust, production-ready automation across diverse enterprise workloads.

ABSTRACT

Autonomous agents driven by Large Language Models (LLMs) offer enormous potential for automation. Early proof of this technology can be found in various demonstrations of agents solving complex tasks, interacting with external systems to augment their knowledge, and triggering actions. In particular, workflows involving multiple agents solving complex tasks in a collaborative fashion exemplify their capacity to operate in less strict and less well-defined environments. Thus, a multi-agent approach has great potential for serving as a backbone in many industrial applications, ranging from complex knowledge retrieval systems to next generation robotic process automation. Given the reasoning abilities within the current generation of LLMs, complex processes require a multi-step approach that includes a plan of well-defined and modular tasks. Depending on the level of complexity, these tasks can be executed either by a single agent or a group of agents. In this work, we focus on designing a flexible agent engineering framework with careful attention to planning and execution, capable of handling complex use case applications across various domains. The proposed framework provides reliability in industrial applications and presents techniques to ensure a scalable, flexible, and collaborative workflow for multiple autonomous agents working together towards solving tasks.

Motivation & Objective

• To address the limitations of single-agent LLMs in industrial automation by enabling collaborative, multi-agent task execution.
• To design a flexible, extensible framework that integrates with existing IT ecosystems, including legacy systems.
• To improve reliability and scalability of AI-driven workflows through structured planning, tool use, and error recovery mechanisms.
• To incorporate human feedback and episodic memory for enhanced learning and stateful task execution.
• To enable production-grade automation by supporting restartable, modular, and auditable workflows.

Proposed method

• The framework employs a modular Agent Unit composed of specialized agents (e.g., Planner, Coder, Tester, Critic) with distinct roles and tools.
• It uses a hybrid agent selection strategy combining semantic matching and mention-based message routing to dynamically assign tasks.
• A Conversational Plan-Reason-Act (ConvPlanReAct) prompting strategy enables structured, multi-turn dialogues among agents.
• Tools are refined at each step using implicit input/output schemas to ensure accurate and reliable function calls.
• Episodic memory stores and retrieves past task results, enabling context-aware reasoning and reuse across plans.
• The system supports intentional and incidental human feedback, with restart capabilities using memory to resume from the last valid state.

Figure 1: Generic agent workflow starting with user input and ending with providing workflow output. Agent Workflow highlights major components and levels of the workflow with (1) Planning, (2) Execution, and (3) Verification done by dedicated agents.

Experimental results

Research questions

• RQ1 How can multi-agent collaboration improve the reliability and scalability of LLM-based task automation in industrial settings?
• RQ2 What architectural patterns enable seamless integration of AI agents with existing enterprise IT ecosystems?
• RQ3 How can episodic memory and structured prompting enhance agent reasoning and task continuity?
• RQ4 What mechanisms ensure robust tool usage and error recovery in complex, multi-step workflows?
• RQ5 How can human feedback be effectively integrated into autonomous agent workflows to improve outcome quality?

Key findings

• The framework enables stable, production-ready automation by supporting restartable workflows using episodic memory to resume from the last valid task.
• The ConvPlanReAct prompting strategy successfully extends ReAct with multi-agent dialog, improving coordination and task decomposition.
• Tool usage is enhanced through dynamic refinement and schema-based input/output validation, increasing reliability and accuracy.
• Episodic memory enables agents to access and reuse results from prior tasks, improving efficiency and consistency across plans.
• Human feedback—both intentional and incidental—is effectively incorporated into the workflow, enhancing final output quality and adaptability.
• The framework supports modular, domain-agnostic workflows across diverse use cases, including document editing and software development, demonstrating broad industrial applicability.

Figure 2: Components contributing to initialization of an Agent. Dashed lines indicate optional components.