Architecture direction for DMAS-Forge: integrated capabilities vs abstraction layer

Determine whether DMAS-Forge should implement comprehensive agentic capabilities internally within the framework or function as an abstraction layer over existing agentic programming frameworks (e.g., LangGraph, CrewAI, AutoGen, LlamaIndex, Agno), thereby decoupling multi-agent workflow logic from protocol-compliant distributed communication and deployment infrastructure.

Background

DMAS-Forge is proposed as a compiler-based framework to decouple multi-agent application logic from communication protocols and deployment infrastructure, enabling transparent generation of glue code and configurations for distributed multi-agent systems across diverse environments.

Current agentic frameworks typically hard-code monolithic communication primitives, complicating distributed deployment and protocol compliance. The paper highlights a fundamental design decision for DMAS-Forge: either to build full agentic capabilities directly within the framework or to act as an abstraction layer atop existing frameworks. The latter would entail techniques (such as monkey patching) to redirect framework-specific communication primitives to standards-compliant channels (e.g., A2A, ACP, MCP) while minimizing re-engineering.