AGENTS.md Is Dead. Meet Its Successor.

March 29, 2026 · 4 min read

You've seen it. Every AI coding repo has one. A single AGENTS.md file trying to explain the entire codebase to Claude or Cursor.

It works for small projects. Then your app grows. The AGENTS.md becomes a novel. The AI starts hallucinating context it should be able to look up. Key decisions get buried under 500 lines of outdated docs.

It's broken and everyone knows it.

Enter lat.md

Agent Lattice just dropped on Hacker News today and it's exactly what vibe coders need.

The concept is simple: replace your monolithic AGENTS.md with a knowledge graph — a folder of interconnected markdown files that live in lat.md/ at your project root.

Each file covers one domain. Auth logic. Architecture decisions. Test specs. They link to each other with [[wiki links]] and back to your actual source code with // @lat: comments.

Why This Actually Works

The problem with AGENTS.md isn't format. It's scale.

When your project has 10 components and 50 API endpoints, no single document can keep it all straight. Your AI assistant ends up guessing. That's where bugs come from.

lat.md solves this with:

• Cross-references — [[cli#search#Indexing]] connects concepts into a navigable graph
• Source code linking — your actual files point back to documentation sections
• lat check — validates that nothing drifts out of sync (this is the killer feature)
• Semantic search — ask "how do we auth?" and get the right section, not a guess
• MCP server — plugs straight into your editor or AI agent

The Setup Takes 2 Minutes

npm install -g lat.md
lat init

That scaffolds your lat.md/ directory. Then you start writing markdown files describing your architecture, business logic, whatever matters.

Your project ends up looking like this:

my-project/
├── lat.md/
│   ├── architecture.md
│   ├── auth.md
│   └── tests.md
├── src/
│   ├── auth.ts      # // @lat: [[auth#OAuth Flow]]
│   └── server.ts    # // @lat: [[architecture#Request Pipeline]]
└── ...

Your AI coding agent can now query exactly the context it needs. No more guessing. No more hallucinating.

Who Built This?

The repo is by 1st1 — Yury Selivanov, the co-creator of Python's asyncio. The dude knows async systems and developer tooling at a deep level.

It's MIT licensed, already has 24 stars on GitHub (just posted today), and requires Node.js 22+. Semantic search needs an OpenAI or Vercel AI Gateway key.

Why This Matters for Vibe Coders

Here's the thing. If you're building with AI coding tools (and you should be), your bottleneck isn't writing code. It's context.

The better your AI understands your codebase, the better the code it writes. lat.md gives you a structured way to feed that context without maintaining a wall of text that goes stale in a week.

I'm adding this to my workflow immediately. You should too.

Want more tools like this? I find the best AI dev tools so you don't have to. Check out the OpenClaw Ultimate Setup — everything you need to build with AI agents, $29 one-time.