← agentic systems Under the Hood — a Single Co-authoring Turn The Hive

Visual explainer of the five-layer architecture beneath a chat with Claude.

Under the hood: a single co-authoring turn
Five layers stack to make a session like this one possible. The substrate at the bottom is the autoregressive next-token prediction described in the transformer literature. What you experience as Claude "thinking" is the upper layers composing on top of that substrate. Click any layer for the concept, a human analogy, and a concrete example.
What is emergent here. The higher you go in the stack, the less it looks like "one model predicting tokens" and the more it looks like distributed cognition. Layer 4 dispatches specialists; Layer 5 integrates them with persistent memory and with the human in the loop. The substrate is still token prediction — but the qualitative behavior shifts character at each layer up. The system approximates collective rather than individual reasoning.
Saved 2026-05-07 from a Cowork chat session — Brian Skene + Claude, exploring the layered architecture beneath agentic AI.