Layer Model
The reference architecture is organized into seven layers. Each layer can evolve independently, but the layers should remain easy to compose.
Use the Architecture Overview for the canonical diagram and Technology Stack for common implementation technologies.
| Layer | Name | Purpose | Initial Examples |
|---|---|---|---|
| 7 | Applications | User-facing and organization-facing agent workflows | Research agents, coding agents, enterprise agents |
| 6 | Models | Hosted and local language model providers | OpenAI, Gemini, Claude, Ollama, llama.cpp, vLLM |
| 5 | Memory | Long-term, semantic, and structured knowledge | Vector databases, knowledge graphs |
| 4 | Agent Protocols | Agent-to-tool and agent-to-agent interoperability | MCP, A2A |
| 3 | APIs | Deterministic application and system interfaces | REST, GraphQL, OpenAPI |
| 2 | Identity | Authentication, authorization, delegation, and policy | OAuth2, OpenID Connect, JWT |
| 1 | Discovery | Machine-readable service and agent discovery | Agent Cards, robots.txt, sitemap.xml, OpenAPI |
Layer Responsibilities
Layer 1: Discovery
Agents need predictable ways to find services, understand capabilities, and locate machine-readable metadata.
Layer 2: Identity
Agents need secure ways to act on behalf of users, organizations, or service accounts while preserving permissions and auditability.
Layer 3: APIs
Agents need stable system interfaces that can be inspected, called, tested, and governed.
Layer 4: Agent Protocols
Agents need protocols for tools, resources, prompts, capability exchange, task delegation, and multi-agent collaboration.
Layer 5: Memory
Agents need durable retrieval systems for semantic search, RAG, entity relationships, and historical context.
Layer 6: Models
Agents need model access through provider-neutral abstractions that allow hosted and local models to be swapped or combined.
Layer 7: Applications
Applications compose the lower layers into useful workflows for research, software engineering, operations, and enterprise automation.