AgentPave

The SpringBoot for AI Agents — a simple, standardised, and extensible framework for building agents that run on any runtime, any cloud, any model.

What is AgentPave?

AgentPave is a concept-first, runtime-agnostic framework that gives developers a simple, standardised, and extensible way to create, run, and manage AI agents.

You define your agent once. AgentPave runs it on LangGraph or MAF — your choice.

[ Your Agent Definition ]
         ↓
    [ AgentPave ]
     ↓         ↓
LangGraph     MAF        ← you choose the runtime
     ↓         ↓
 Any LLM    Any LLM      ← model agnostic
     ↓         ↓
Any Cloud  Any Cloud     ← cloud agnostic

Why AgentPave?

Every agent framework today makes you choose a runtime and live with its opinions. LangGraph is powerful but complex. MAF is enterprise-grade but Azure-first. CrewAI is fast but hits walls in production.

AgentPave sits above all of them. It gives you:

• A clean agent definition model — declare what your agent is, not how it runs
• Runtime portability — swap LangGraph for MAF without touching your agent code
• Production-ready from day one — memory, reliability, security, cost management, and observability built into the spec
• Extensible by design — start with core, add capabilities as you grow
• Developer-first — running your first agent in under 5 minutes

Core Principles

What AgentPave Covers

AgentPave is built around 13 dimensions — the complete set of concerns every production agent framework must address.

Reference Implementations

The same agent definition runs on both. If it doesn't — the spec has a gap.

Extensions

AgentPave is extensible by design. Official extensions plug into defined extension points.

What AgentPave is NOT

• Not a cloud platform — no managed infrastructure
• Not a UI or drag-and-drop builder
• Not tied to any LLM provider
• Not tied to any cloud provider
• Not a replacement for LangGraph or MAF — it sits above them

Documentation

Status

Contributing

AgentPave is concept-first. The most valuable contributions right now are:

• Feedback on the Baseline Structure
• Gaps you've experienced building agents in production that the spec doesn't address
• Reference implementation contributions (LangGraph, MAF, and beyond)

AgentPave — Build agents, not plumbing.