agentpave 0.1.0

The runtime-agnostic AI agent specification. Define once. Run anywhere.

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

Principle	What it means
Progressive Autonomy	Start with human involvement. Reduce it as the agent proves itself.
Durability over Convenience	Agents survive failures and resume — never restart from scratch.
Determinism where Possible	LLM for reasoning. Typed code for execution.
Version Everything	Agents, LLM interfaces, schemas, tool contracts — all versioned.
Least Privilege Always	Agents never exceed the privilege of the task that invoked them.
Cost Awareness by Default	Every agent declares a budget. No agent runs unconstrained.
Language & Runtime Agnostic	The spec has no runtime opinions. Python is first — not only.

---

What AgentPave Covers

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

#	Dimension	What it solves
1	Identity	What the agent is and how it is declared
2	Lifecycle	Declared → Registered → Running → Paused → Resumed → Retired
3	Communication	MCP for tools, A2A for agents, defined human communication model
4	Memory	Working, Episodic, Semantic (RAG), Procedural — with lifecycle
5	Reliability	Durable execution, drift detection, reliability contracts
6	Security	Zero Trust identity, privilege model, policy-as-code
7	Economics	Token budgets, model routing, prompt caching
8	Observability	OpenTelemetry hooks, step-level tracing, pluggable backends
9	Testability	CI/CD integration, evaluation framework, production feedback loop
10	Developer Experience	CLI, local runner, mock tools, templates, 5-minute hello world
11	Extensibility	Core never depends on extensions. Extensions always depend on core.
12	Governance	Policy declaration, human oversight hooks, compliance adapters
13	Portability	No lock-in at any layer — runtime, model, cloud, or language

---

Reference Implementations

Implementation	Runtime	Language
AgentPave-LangGraph	LangGraph	Python
AgentPave-MAF	Microsoft Agent Framework	Python + .NET

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.

Extension	Purpose	Priority
AgentPave-Observe	OpenTelemetry backend integration	Release 1
AgentPave-Eval	Evaluation + CI/CD + production feedback loop	Release 1
AgentPave-HITL	Human-in-the-loop approval workflows	Release 1
AgentPave-MultiAgent	Agent-to-agent coordination	Release 2
AgentPave-Cache	Semantic + prompt caching	Release 2
AgentPave-Memory	Memory consolidation	Release 2
AgentPave-Govern	EU AI Act, SOC2, HIPAA, GDPR adapters	Release 3
AgentPave-Marketplace	Agent registry and discovery	Release 3

---

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

Document	Description
Baseline Structure	The complete, versioned AgentPave structure — the north star document
AgentPave Spec v1.1.2	Master specification — 13 dimensions, 86 acceptance criteria, 32-error taxonomy
D1 — Identity	Agent identity, SPIFFE, AgentRegistry, Runtime Tokens
D2 — Lifecycle	6 lifecycle states, hooks, audit log
D3 — Communication	MCP, Integration Contracts, circuit breaker, A2A
D4 — Memory	Working, Episodic, Semantic, Procedural memory
D5 — Observability	OpenTelemetry, AgentResult schema, pluggable backends
D6 — Reliability	Durable execution, checkpoints, deterministic routing, drift detection
D7 — Security	Zero Trust, OPA/Cedar policy, OWASP Agentic Top 10
D8 — Economics	Token budgets, model routing, semantic caching
D9 — Testability	Mock tools, evaluation framework, CI/CD gates
D10 — Developer Experience	CLI, local runner, 5-minute hello world
D11 — Extensibility	8 declared extension points, manifest schema
D12 — Governance	EU AI Act, SOC2, HIPAA, GDPR compliance adapters
D13 — Portability	Interoperability test suite, RuntimeAdapter interface
AgentPave-LangGraph	(coming soon) Reference implementation — LangGraph runtime
AgentPave-MAF	(coming soon) Reference implementation — Microsoft Agent Framework

---

Status

Item	Status
Baseline Structure	✅ v1.0 — Baselined June 2026
AgentPave Spec	✅ v1.1.2 — Published June 2026
MVP Dimensions (D1–D6)	✅ Complete — 51 acceptance criteria
Post-MVP Dimensions (D7–D13)	✅ Complete — 35 acceptance criteria
Error Taxonomy	✅ Complete — 32 error types
AgentPave-LangGraph	📅 Planned
AgentPave-MAF	📅 Planned

---

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.