FIELD MAP
THE AI STACK — MAY 2026
The application layer most teams ship on is now ten distinct layers deep, wrapped by two rails that touch every one of them. This is a working map, not a buyer's guide: where each category sits, what a few representative providers do, and how the pieces connect.
Read it top to bottom — surface to silicon. The left rail, Observability, and the right rail, Governance, are not steps in the flow; they are concerns that cut across all ten layers. Tap any provider in the diagram to jump to its explanation and an outbound link below.
- 01
End-User Surfaces
- 02
Agent Runtimes
- 03
Orchestration Frameworks
- 04
Protocol Layer NEW
- 05
Memory NEW
- 06
Retrieval
- 07
Storage
- 08
Model Gateway
- 09
Foundation Models
- 10
Inference + Compute
Tap any item for details ↓
01End-User Surfaces
Cursor
AI-first code editor; agentic edits and codebase-wide changes from natural language.
Visit providerPerplexity
Answer engine: conversational search with live sources and citations.
Visit providerChatGPT
OpenAI's consumer assistant for chat, reasoning and tool use.
Visit providerClaude
Anthropic's assistant across web, desktop and mobile, tuned for long-context work.
Visit provider02Agent Runtimes
Claude Code
Terminal-native agentic coding from Anthropic; delegates multi-step engineering tasks.
Visit providerDevin
Cognition's autonomous software engineer that plans and executes end-to-end.
Visit providerReplit Agent
Builds and deploys full apps from a prompt inside Replit's cloud IDE.
Visit providerCodex
OpenAI's coding agent for the cloud and CLI, running tasks in isolated sandboxes.
Visit providerCursor Agent
Cursor's background agent mode for parallel, longer-running coding work.
Visit provider03Orchestration Frameworks
LangGraph
Graph-based orchestration for stateful, multi-step agent workflows (LangChain).
Visit providerMicrosoft Agent Framework
Microsoft's unified agent framework, consolidating Semantic Kernel and AutoGen.
Visit providerPydantic AI
Type-safe Python agent framework from the Pydantic team.
Visit providerMastra
TypeScript framework bundling agents, workflows, memory and evals.
Visit providerGoogle ADK
Google's open-source Agent Development Kit (Python, Java, Go, TypeScript).
Visit provider04Protocol Layer
MCP
Model Context Protocol (Anthropic): a standard way to connect models to tools and data.
Visit providerA2A
Agent2Agent: cross-vendor agent interoperability; created by Google, now Linux Foundation.
Visit providerAG-UI
Agent-User Interaction protocol (CopilotKit): event stream between agent backends and frontends.
Visit provider05Memory
Mem0
Drop-in memory API combining vector, graph and key-value stores for personalization.
Visit providerLetta
OS-style agent memory with paging between context and archival storage (formerly MemGPT).
Visit providerZep
Temporal knowledge-graph memory (Graphiti) that tracks how facts change over time.
Visit provider06Retrieval
Cohere Rerank
Reranking models that reorder candidate passages by true relevance.
Visit providerVoyage AI
High-quality embedding and reranking models (part of MongoDB).
Visit providerNeo4j GraphRAG
Graph-based RAG that grounds retrieval in a knowledge graph.
Visit providerElastic
Hybrid keyword and vector search on the Elasticsearch engine.
Visit provider07Storage
pgvector
Postgres extension adding vector similarity search to an existing database.
Visit providerQdrant
Open-source vector database with payload filtering and hybrid search.
Visit providerTurbopuffer
Serverless vector and full-text search built on object storage for low cost at scale.
Visit providerPinecone
Fully managed vector database for production retrieval.
Visit providerneo4j
Native graph database for richly connected data.
Visit provider08Model Gateway
Portkey
AI gateway adding routing, caching, guardrails and observability across providers.
Visit providerLiteLLM
Unified SDK and proxy exposing 100+ model providers behind one OpenAI-style API.
Visit providerOpenRouter
A single API that routes requests across many models and providers.
Visit provider09Foundation Models
Claude (Anthropic)
Anthropic's Claude model family, tuned for reasoning, coding and long context.
Visit providerGPT (OpenAI)
OpenAI's GPT family of general-purpose frontier models.
Visit providerGemini (Google)
Google DeepMind's multimodal Gemini model family.
Visit providerMeta (Llama)
Meta's open-weight Llama models for self-hosting and fine-tuning.
Visit providerDeepSeek
Open-weight models known for strong reasoning at low cost.
Visit providerQwen
Alibaba's open-weight Qwen model family across sizes and modalities.
Visit provider10Inference + Compute
Together AI
Inference cloud for running and fine-tuning open models at scale.
Visit providerFireworks AI
Fast, cost-efficient inference serving for open models.
Visit providervLLM
Open-source high-throughput inference engine for LLM serving.
Visit providerNVIDIA
Data-center GPUs that dominate AI training and inference.
Visit providerAMD MI400
AMD's Instinct MI400-series AI accelerators — AMD's datacenter challenge to NVIDIA.
Visit providerGoogle TPU
Google's Tensor Processing Units for training and serving on Google Cloud.
Visit providerAWS
Cloud plus custom Trainium and Inferentia silicon for AI workloads.
Visit providerGroq
LPU-based inference delivering very low-latency token generation.
Visit provider