# The AI-Native Stack

**Track:** Building the AI-Native Product — AI for Entrepreneurship — complete (29)
**Framework / surface:** venture strategy
**Level:** Intermediate
**Prerequisites:** Unit Economics of Inference

**In one line:** Models, retrieval, orchestration, evals — what to build, what to rent, what to expect to throw away.

## Theory, aesthetics & inspiration

The stack above the model has stabilized into recognizable layers: retrieval that grounds generation in your data — the pattern formalized as retrieval-augmented generation by Patrick Lewis and colleagues in 2020 — orchestration that sequences model calls and tools, evaluation harnesses that say whether any of it works, and the models themselves, rented from labs or run as open weights. The build-versus-rent line keeps moving as capability compounds, and the durable rule is to own what differentiates and rent what commoditizes: your evals, your data pipelines, and your workflow integration are yours; the model layer is a supplier relationship with unusually healthy vendor competition. Anthropic's Model Context Protocol (2024) marks the architectural turn worth building toward — a standard interface for connecting models to tools and data, on the premise that software increasingly presents itself to agents, not only to people. Expect to rewrite a large fraction of this stack yearly; design so the rewrites are cheap.

**Founder question:** Which layer of your stack would hurt a competitor most to lose — and do you own it?