The AI Agent Stack Map
Every layer of a modern AI agent system, mapped. Eight layers, from the foundation model at the bottom to deployment at the top. This page is updated quarterly and designed to be the canonical reference you can link to when explaining the agent stack.
1. Foundation Models
The LLM that powers reasoning and generation.
Examples
- Claude Opus 4.7 — Anthropic
- GPT-5 — OpenAI
- Gemini 2.5 — Google
- Llama 4 — Meta
- DeepSeek R1 — DeepSeek
2. Agent Framework
Orchestration libraries that coordinate model calls, tool use, memory, and multi-agent collaboration.
Examples
- Claude Agent SDK — Anthropic
- LangGraph — LangChain
- CrewAI — CrewAI
- AutoGen — Microsoft
- Vercel AI SDK — Vercel
3. Tool Use Layer
The protocol + runtime that lets agents call external tools. MCP (Model Context Protocol) is the emerging standard.
Examples
- MCP — Anthropic (open)
- OpenAI Function Calling — OpenAI
- Custom tool schemas — Various
4. Skills & Capabilities
Packaged agent capabilities — browsing, code execution, file ops, memory, vector search — installable as skill bundles.
Examples
- Web Browsing — agentmag
- Code Execution — agentmag
- Memory & Retrieval — agentmag
- Vector Search — agentmag
- Terminal — agentmag
5. Memory Layer
Short-term context plus long-term memory across sessions.
Examples
- Cosmos DB / Mongo — Azure / MongoDB
- Pinecone — Pinecone
- Weaviate — Weaviate
- pgvector — Postgres
6. Evaluation Layer
How you measure agent quality — task success rate, cost per task, latency, safety.
Examples
- Braintrust — Braintrust
- LangSmith — LangChain
- Custom harnesses — Various
7. Observability
Tracing and logging of agent executions — critical for debugging non-deterministic failures.
Examples
- LangSmith — LangChain
- Braintrust — Braintrust
- Sentry — Sentry
- OpenTelemetry — CNCF
8. Deployment
How you run agents in production — serverless, dedicated infra, or vendor platforms.
Examples
- Azure App Service — Microsoft
- Vercel — Vercel
- AWS Bedrock Agents — AWS
- Modal — Modal
- E2B — E2B
Frequently Asked
What is the AI agent stack?
The AI agent stack is the set of layers required to build a production AI agent: a foundation model, an orchestration framework, a tool use layer (often MCP), installable skills, a memory layer, evaluation harnesses, observability, and deployment infrastructure. Each layer has multiple competing vendors and open-source options.
What's different about the agent stack vs a traditional ML stack?
Traditional ML stacks are dominated by training and serving. Agent stacks are dominated by orchestration: the model is commodity; the hard problem is tool use reliability, multi-step reasoning, memory, and evaluation under non-deterministic outputs. You'll spend more time on the framework, skills, and observability layers than on the model itself.
Do I need all 8 layers?
No. A minimal production agent can run with just a foundation model + a light framework + one or two tools. The full stack becomes necessary as you scale: multi-step reasoning needs observability, persistent context needs memory, quality control needs evaluation. Start with layers 1–3 and add layers as bottlenecks appear.
Is the agent stack stable in 2026?
The foundation model, deployment, and observability layers are stable. The framework, tool use, and skills layers are still churning — expect consolidation over the next 12–18 months. MCP is emerging as a standard for tool use, which should stabilize layer 3 specifically.