Design Docs Are All You Need: DeepMind's library regenerates all code from NL docs

omarsar0 · x · 2026-09-07

A strange new paper from Google DeepMind, MIT and colleagues maintains an ML performance-modeling library whose main branch contains almost no code — instead, the repo is a directed graph of natural-language design docs, and coding sub-agents regenerate the full implementation on every version bump. Every human change is a doc edit.

The premise: ML performance modeling invalidates its own abstractions every hardware/model generation, and coding agents are now cheap enough that regenerating beats patching. Two things make regeneration reliable: docs are written around step-by-step worked examples that act as in-context demonstrations, and the system is anchored on a minimal recursive core. A radical experiment in docs-first, code-as-derived-artifact software engineering for the agent era.

Original post →

Design Docs Are All You Need: DeepMind's library regenerates all code from NL docs

More from coding & agent