Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Behavior Modeling for Training-free Building of Private Domain Multi Agent System

Won Ik Cho, Woonghee Han, Kyung Seo Ki, Young Min Kim

TL;DR

The paper tackles the difficulty of deploying LLM-driven, tool-using agents in private domains where tools are heterogeneous, jargon is domain-specific, and governance is strict. It introduces a training-free framework based on behavior modeling and SpecDoc, enabling an orchestrator, a tool-calling agent, and a general chat agent to work with private tools without fine-tuning. By externalizing domain knowledge in structured SpecDoc documents, the approach preserves general model capabilities while achieving domain alignment, supported by a development workflow, versioning, and live knowledge retrieval aspects. The work presents use cases like zero-shot deployment, synthetic dialogue generation for evaluation, and SpecDoc as a living knowledge base, arguing for a scalable, transparent path to vertical AI systems in enterprises. Overall, this framework offers a sustainable, maintainable alternative to retraining large models for each private domain, reducing drift and aligning behavior with domain expertise in private conversational ecosystems.

Abstract

The rise of agentic systems that combine orchestration, tool use, and conversational capabilities, has been more visible by the recent advent of large language models (LLMs). While open-domain frameworks exist, applying them in private domains remains difficult due to heterogeneous tool formats, domain-specific jargon, restricted accessibility of APIs, and complex governance. Conventional solutions, such as fine-tuning on synthetic dialogue data, are burdensome and brittle under domain shifts, and risk degrading general performance. In this light, we introduce a framework for private-domain multi-agent conversational systems that avoids training and data generation by adopting behavior modeling and documentation. Our design simply assumes an orchestrator, a tool-calling agent, and a general chat agent, with tool integration defined through structured specifications and domain-informed instructions. This approach enables scalable adaptation to private tools and evolving contexts without continual retraining. The framework supports practical use cases, including lightweight deployment of multi-agent systems, leveraging API specifications as retrieval resources, and generating synthetic dialogue for evaluation -- providing a sustainable method for aligning agent behavior with domain expertise in private conversational ecosystems.

Behavior Modeling for Training-free Building of Private Domain Multi Agent System

TL;DR

The paper tackles the difficulty of deploying LLM-driven, tool-using agents in private domains where tools are heterogeneous, jargon is domain-specific, and governance is strict. It introduces a training-free framework based on behavior modeling and SpecDoc, enabling an orchestrator, a tool-calling agent, and a general chat agent to work with private tools without fine-tuning. By externalizing domain knowledge in structured SpecDoc documents, the approach preserves general model capabilities while achieving domain alignment, supported by a development workflow, versioning, and live knowledge retrieval aspects. The work presents use cases like zero-shot deployment, synthetic dialogue generation for evaluation, and SpecDoc as a living knowledge base, arguing for a scalable, transparent path to vertical AI systems in enterprises. Overall, this framework offers a sustainable, maintainable alternative to retraining large models for each private domain, reducing drift and aligning behavior with domain expertise in private conversational ecosystems.

Abstract

The rise of agentic systems that combine orchestration, tool use, and conversational capabilities, has been more visible by the recent advent of large language models (LLMs). While open-domain frameworks exist, applying them in private domains remains difficult due to heterogeneous tool formats, domain-specific jargon, restricted accessibility of APIs, and complex governance. Conventional solutions, such as fine-tuning on synthetic dialogue data, are burdensome and brittle under domain shifts, and risk degrading general performance. In this light, we introduce a framework for private-domain multi-agent conversational systems that avoids training and data generation by adopting behavior modeling and documentation. Our design simply assumes an orchestrator, a tool-calling agent, and a general chat agent, with tool integration defined through structured specifications and domain-informed instructions. This approach enables scalable adaptation to private tools and evolving contexts without continual retraining. The framework supports practical use cases, including lightweight deployment of multi-agent systems, leveraging API specifications as retrieval resources, and generating synthetic dialogue for evaluation -- providing a sustainable method for aligning agent behavior with domain expertise in private conversational ecosystems.

Paper Structure

This paper contains 15 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Agents and Multi-Agent Systems in LLM Era
  4. LLM Adaptation: Fine-Tuning vs. Prompting
  5. Challenges in Specialized Domains
  6. Multi-Agent System for Private Domains
  7. Behavior Modeling and Data Abstraction
  8. Incorporating SpecDoc into Agents' Prompts
  9. Development and Maintenance of Behavior Modeled MAS
  10. Comparison with Traditional Approaches
  11. Use Cases
  12. Synthetic Dialogue Generation
  13. SpecDoc as a Living Knowledge Base
  14. Zero-Shot Execution of Multi-Agent Systems
  15. Discussion and Conclusion

Figures (1)

  • Figure 1: Development and maintenance flow of the behavior-modeled multi-agent system. SpecDoc is created from domain tool definitions and expert scenarios and fed into the instructions of each agent and API docstrings, and iteratively refined after evaluation of domain experts. Updates to tools or domain knowledge are handled by editing SpecDoc and updating agent prompts accordingly, with version control. The illustration was assisted by https://app.eraser.io/.