AgentMesh
A reference architecture for an enterprise AI platform: provider-agnostic LLM access, agent runtime, RAG and MCP behind a hexagonal core.
Overview
AgentMesh is a portfolio reference architecture under active development. It is not established production experience; it exists to demonstrate how a reusable enterprise AI platform can be structured.
Problem
AI prototypes often hard-wire one LLM provider, one vector store and one orchestration library, which makes them difficult to evolve, test or move between clouds.
Solution
Use Clean / Hexagonal architecture so LLM providers, agent implementations, RAG providers, tools and infrastructure adapters are all replaceable behind ports.
Architecture
- 01
Interface
- FastAPI
- Pydantic models
- 02
Application
- Agent runtime
- Workflow orchestration
- 03
Ports
- LLM provider
- RAG provider
- Tools / MCP
- 04
Adapters
- OpenAI / Azure OpenAI / local LLMs
- PostgreSQL
- Redis
- RabbitMQ
- 05
Operations
- Docker
- Kubernetes
- Observability
- CI/CD
- Ports and adapters for LLM providers: OpenAI, Azure OpenAI and local models.
- Async Python and Pydantic models throughout the API and runtime.
- LangGraph, LlamaIndex, A2A and RabbitMQ are project and learning implementations, not claimed production experience.
- Observability, evaluation and CI/CD are part of the design from the start.
Key features
- Replaceable LLM providers, agents, RAG providers and tools
- Hexagonal core with infrastructure adapters
- Agent runtime with tool and MCP integration
- Designed for observability and evaluation
Engineering challenges
- Keeping the core free of provider-specific SDK details.
- Testing agent behaviour deterministically.
- Choosing how much orchestration framework to adopt versus own.
Results
No measured outcomes are published for this project. I only list results that have been verified.