Towards Declarative Composition of Heterogeneous AI Agents

What is it about?

Building AI applications with multiple agents can be difficult because each agent may use different tools, models, and technical settings. Our research proposes a simpler way to connect AI agents. Instead of writing custom integration code for each agent, developers describe them using a standard configuration document called an **Agent Manifest**. This document defines what the agent does, how to communicate with it, what information it accepts and returns, and which operational rules it must follow. We tested this approach with five AI agents that analyze Java code, showing that different agent roles can use the same execution system. The goal is to make AI agents easier to integrate, manage, and change.

Featured Image

Photo by Deng Xiang on Unsplash

Photo by Deng Xiang on Unsplash

Why is it important?

Many AI agent solutions depend on specific frameworks, models, or providers. This can make applications harder to maintain, update, and reuse. Our research explores a more flexible approach: separating the description of an AI agent from the technology used to run it. This allows developers to manage agents through reusable configuration documents while using common rules to validate their inputs, outputs, and operational limits. In our experiments, five agents shared the same execution process. The system also detected invalid data, identified operational limit violations, and allowed an AI model to be replaced by changing only one configuration field. These initial results provide a foundation for building AI systems that are more portable, easier to monitor, and less dependent on specific technologies.

Perspectives