From Agents to Teams: Optimizing the Creation of LLM-based Multi-Agent Systems

As multi-agent systems become more complex, improving how Large Language Model (LLM)-based agents work together is increasingly important. This research focuses on creating reusable design patterns and adaptive coordination strategies to boost collaboration and system performance. Blend360 is an AI service provider that builds advanced solutions for clients in sectors like finance, healthcare, and consumer goods. The company is exploring the use of multi-agent systems in its AI toolkit. These systems break down client problems into smaller tasks that multiple AI agents can solve together. However, there are key challenges preventing their smooth deployment in real-world applications. First, current systems often use static workflows (such as sequential or hierarchical setups), which limit flexibility in dynamic environments. Second, there is a lack of reusable design patterns to guide how agents interact, making these systems hard to build and maintain. Third, existing evaluation methods focus only on whether a task is completed, ignoring how well agents collaborate, how much each contributes, or how efficient their communication is. This research will help address these issues and directly support the development of smarter, more scalable agentic AI tools for Blend360s clients.

Faculty Supervisor:

An Ran Chen

Student:

Partner:

Blend360

Discipline:

Engineering

Sector:

Professional, scientific and technical services

University:

University of Alberta

Program: