← Architectures

The agentic AI stack

Nine layers grouped into three planes: Data (what the agent knows about), Control (how it decides and acts), and Governance (how it stays safe and correct). Click any layer to open it up — sub-types, real tools, the papers to read, and animated lines showing which other layers it directly touches.

How to read this

The 3D scene is the entry point, not the whole story. Every layer is a discipline of its own — memory alone has three or four sub-types with distinct failure modes, evaluation has six ways to be done, security has a whole OWASP top-10. The panel on the right is where the real content lives. When you click a layer:

• Why it matters and what breaks without it — the argument for the layer's existence, first.
• Sub-types and patterns — the internal shape. RAG isn't just "retrieval"; it's dense vs. sparse vs. hybrid vs. reranking vs. GraphRAG.
• In the wild — real tools and vendors, not generic categories.
• Read more — canonical links out. Papers, docs, primers. Follow these when a layer catches your attention.
• Directly interacts with — the specific relationships to other layers, plus animated flow lines in the scene showing them.

What's new here vs. the usual diagram

Most agent-stack diagrams online lump memory into "data" and lump tools into the agent itself. In production those are separate concerns with their own failure modes, so they're broken out here. Model gateway is the third addition — you don't feel the need for it until you've shipped a system that hard-coded one provider and then had to migrate.

Everything visible here maps to references you can go read for yourself: Anthropic's Building Effective Agents, LangChain's What is an Agent?, and a16z's Emerging Architectures for LLM Applications.

Where to start clicking

If you're new to agent systems, start with Agentic in the Control plane — it touches everything, so you'll see the whole graph light up. If you're already building, start with the layer you know the least about — Model gateway and Memory are the most under-appreciated in most orgs.