Code for AI

The home of AI-native open source

Code should be ready for AI contributors.

Open source was organized for people to discover, understand, modify, and maintain. AI agents are becoming active software contributors. Code for AI defines how software should be published for both.

Definition

AI-native open source

AI-native open source is software published not only for human use and collaboration, but for AI agents to discover, understand, operate, modify, verify, attribute, and govern under explicit, machine-readable constraints.

Human readability remains essential. AI-native software adds a machine-actionable publication layer; it does not replace human inspection, judgment, or accountability.

Why now

The primary software consumer is changing.

A traditional repository assumes that a person will read its documentation, infer its architecture, locate its extension points, interpret its license, and validate every change. Coding agents increasingly perform part of that work.

A README alone cannot reliably communicate every architectural invariant, operational command, modification boundary, verification requirement, attribution rule, and governance constraint an autonomous contributor needs.

Code for AI treats those elements as first-class software assets: explicit, versioned, inspectable, and usable by machines.

Core principles

What makes a repository AI-native?

01

Discoverable

Purpose, capabilities, interfaces, ownership, and canonical sources are explicit.

02

Understandable

Architecture, concepts, dependencies, invariants, and design decisions are represented clearly.

03

Operable

Setup, execution, tool use, permissions, and recovery paths are stable and machine-actionable.

04

Modifiable

Extension points, generated files, compatibility boundaries, and prohibited changes are declared.

05

Verifiable

Builds, tests, benchmarks, security checks, and acceptance criteria can validate agent changes.

06

Attributable

Licensing, authorship, dependencies, provenance, and derivative work remain traceable.

07

Governable

Authority, review requirements, policy boundaries, and human approval points are enforceable.

08

Human-accountable

AI participation never removes the need for human oversight, auditability, and responsibility.

Maturity model

From human-oriented to AI-native.

Level 0

Human-oriented

The repository is documented primarily for people. Agents must infer structure, rules, and validation paths.

Level 1

Agent-ready

The repository provides dedicated agent instructions, reliable commands, a navigable architecture, and explicit constraints.

Level 2

AI-native

Agent comprehension, operation, modification, verification, attribution, and governance are designed into the software lifecycle.

Scope

More than an instruction file.

Files such as AGENTS.md are an important step toward agent-ready repositories. AI-native open source is the broader publishing model around them.

It includes repository semantics, stable interfaces, execution contracts, verification, provenance, licensing, authority, and lifecycle governance. The objective is not merely to help an agent generate code, but to help it change software correctly and responsibly.

About the initiative

Code for AI is an open initiative founded by Silan Hu.

Silan Hu is an AI systems researcher and Computer Science PhD student at the National University of Singapore. His work focuses on AI-native data systems, agent infrastructure, verifiable workflows, and the publication of digital assets for AI discovery and use.

Learn more at silan.tech.