Architecture

The Terraphim Engine: Six Layers, Fifty-Two Crates

How the engine behind this reference architecture is actually decomposed: six layers, single-responsibility crates, and why that boundary discipline is what makes an agent system auditable.

The MCP-Native Agent Architecture

Why Model Context Protocol is the missing link between probabilistic LLMs and production-grade agent orchestration — a reference architecture drawn from Terraphim

Harness Engineering for Self-Improving Agents

Proposed reference pattern for supervised feedback loops that make AI agent behavior measurable and improvable

Announcing Reference Architecture for AI

Capability Map (2020-2022)

The original capability map for building AI products: the AI product itself, core AI/ML capabilities, and enabling capabilities. Illustrated with The Pattern.