Lat.md: Agent Lattice: a knowledge graph for your codebase, written in Markdown | Hacker News Reader

Lat.md: Agent Lattice: a knowledge graph for your codebase, written in Markdown

github.com

90 pointsdoppp62 comments

I was already doing what lat.md does — just badly

"why not just put markdown files in each folder?" I had those files. I also had the answer.

I had lots of `.md` files, custom skills for consistency checks, a test-audit skill for coverage gaps, a code-truth skill that worked backwards from code to documentation. All untethered from the actual code. Nothing connected a design decision in `auth-flow.md` to the function implementing it. Nothing told me when a renamed method made the docs fiction.

So I became the binding. Personally running consistency checks. Running reconciliation. Constantly worrying about drift. It didn't scale — it scaled me.

lat.md replaced that work with three things: docs link into source code and source code comments link back to docs — the connection is explicit, not hopeful. `lat check` enforces referential integrity automatically — the tool worries about consistency so I don't have to. And the knowledge compounds — every session the agent consults the graph and can update it. I've taken to ending conversations with "review this and update [x]" where "x" is some document that I'm working on in the thread. lat.md gets updated for free.

I haven't run rigorous tests. It isn't perfect. There's no planning engine but I get by fine with various *plan.md file. But I'm rolling it out across all my repos because it replaced manual work I was already doing, more reliably than I was doing it.

Not "why do I need a tool for documentation?" but "why was I personally doing the job of a linter?"

I've been doing similar work since Claude code updated their "slash command"(later merge to skills), first 3-4 long content docs, gradually split to modular groups. I designed it for loading the docs based on what agent is actually doing. The maintenance part is honestly not that hard, for me, I created some CI jobs that diffs the docs against the codebase and flags drift handles of it.

The pattern works.

But I keep catching myself spending more time on how to organize context than on what the agent is actually supposed to accomplish.

Feels like the whole space has that problem right now.

Creator of lat.md here. There are two videos with me talking about lat in more detail [1] and less detail [2]. But I'm also working on a blog post exploring lat and its potential, stay tuned.

AMA :)

[1] https://x.com/mitsuhiko/status/2037649308086902989?s=20

[2] https://www.youtube.com/watch?v=gIOtYnI-8_c

What does lat.md do better than a docs/ or specs/ directory you manage yourself?

- makes agent manage it for you in a controlled way

- gives agent tools to search and traverse the graph, better than they can with grep

- ability to have semantic links into your code and back

etc, see the README

I read the README and it wasn't clear what the material benefits of these features were over e.g. having the agent write an organize markdown files with guidance, incorporating README.md with indexes/ToCs that act as lookup tables, linking to code with relative filepath links and incorporating descriptions or additional context where needed.

Not trying to poopoo your project or be dismissive, just curious whether I'm missing something.

I found having smaller structured markdowns in each folder explaining the space and classes within keeps Claude and Codex grounded even in a 10M+ LOC codebase of c/c++

I'm working on making lat hierarchical, e.g.

  - lat.md           # high-level description of the project
  - frontend/lat.md  # frontend-related knowledge
  - backend/lat.md   # details about your backend

It’s a start but it wouldn’t solve my uses cases. I developed my own skill for this. Called morning-routine that does a recursive `ls -R` on all the Claude markdowns. It’s staged in multiple stages so I don’t have to waste as much context if I don’t need to.

Keep going, prompt engineering is fascinating.

Reading this phrase gave me brain damage

> "chalk": "^5.6.2",

security.md ist missing apparently.

good catch. Makes me wonder if we could feed the Agent with a repository of known vulnerability and security best practices to check against and get ride of most deps. Just ask _outloud_ so to speak.

Why would you even need chalk on modern Node.js? It can style natively now.

Good idea. :)

https://github.com/1st1/lat.md/commit/da819ddc9bf8f1a44f67f0...

Fixed

This is one of the things that GitHub Spec Kit solves for me. The specify.plan step launches code exploration agents and builds itself the latest data model, migrations, etc etc. Really reduces the need to document stuff when the agent self discovers codebase needs.

Give Claude sqlite/supabase MCP, GitHub CLI, Linear CLI, Chrome or launch.json and it can really autonomously solve this.

Can you offer any insights how this compares to building an AST or RAG over your codebase? Several projects do that and it autoupdates on changes too. The agent does a wide sweep using AST/RAG search followed by drill down using an LSP. This sped up my search phase by 50%. How Will this project help me?

I think it's a great idea and I'm considering building this in lat too. Code embedding models can definitely speed up grepping further, but they still wouldn't help much when you have a business logic detail encoded across multiple complex files. With lat you'd have it documented in a paragraph of text + a few [[..]] links into your code.

Intriguing concept, especially because I've thought of dabbling into this area.

Can this tool solve these problems with models:

- insisting on using sed and shell redirection instead of File Edit Tools?

- trying to use npx to run commands instead of npm run that's right there in package.json

- forgetting to produce the docs that was asked for in AGENTS.md

- checking and using latest package versions instead of deciding somehow that years old versions are good enough

To my mind these are context problems, where somehow model chooses other information it has over what's in the repository, and what tools it has on call.

The other side of Lat.md, checking diffs models make among other changes, that's hard for me to grasp. I'd need to see it in action. Perhaps a coding session as a stream?

We've been doing this with simple mkdocs for ages. My experience is that rendering the markdown to feel like public docs is important for getting humans to review and take it seriously. Otherwise it goes stale as soon as one dev on the project doesn't care.

I definitely agree with the need for this. There's just too much to put into the agents file to keep from killing your context window right off the bat. Knowledge compression is going to be key.

I saw this a couple of days ago and I've been working on figuring out what the right workflows will be with it.

It's a useful idea: the agents.md torrent of info gets replaced with a thinner shim that tells the agent how to get more data about the system, as well as how to update that.

I suspect there's ways to shrink that context even more.

> I suspect there's ways to shrink that context even more.

Yeah, I'm experimenting with some ideas on that, like adding `lat agent` command to act as a subagent to search through lat and summarize related knowledge without polluting the parent's agent context.

The staleness problem mentioned here is real. For agentic systems, a markdown-based DAG of your codebase is more practical than a traditional graph because agents work within context windows. You can selectively load relevant parts without needing a complex query engine. The key is making updates low-friction -- maybe a pre-commit hook or CI job that refreshes stale nodes.

This is what lat is doing. `lat init` would install hooks to double check that Claude/Codex/OpenCode update lat.md when they finish the work.

So the graph is human-maintained, and agents consume it and `lat check` is supposed to catch broken links and code-spec drift. How do you manage this in a multi-agent setup? Is it still a manual merge+fix conflicts situation? That's where I keep seeing the biggest issues with multi-agent setups

The graph can be human-maintained, but the preferred flow is to let agents update it.

Curious if you've seen a difference between agents finding things through the graph structure vs just vector search over a docs directory. The section constraints and check validation seem like the real quality wins here, wondering how much the wiki link topology adds on top of that.

Docs typically focus on the public api, lat focuses on both public and private impl detail

At that point why not have an obsidian vault in your repo and get the Agent to write to it?

I was thinking the same. Especially now that Obsidian has a CLI to work with the vault.

The one thing I saw in the README is that lat has a format for source files to link back to the lat.md markdown, but I don't see why you couldn't just define an "// obs:wikilink" sort of format in your AGENTS.md

Because lat gives agents more tools and enforces the workflow.

Unlike obsidian, lat allows markdown files link into functions/structs/classes/etc too.

This saves agents time on grepping but also allows you to build better workflows with tests.

Test cases can be described as sections in `lat.md/` and marked with `require-code-mention: true`. Each spec then must be referenced by a `// @lat:` comment in test code. `lat check` flags any spec without a backlink, so you can review and maintain test coverage from the knowledge graph.

This is interesting. I've been using a system-wide Obsidian vault that all my agents use for stuff that's platform-specific instead of project specific (think common Android/Samsung-related or ANR fixes). So far, it hasn't been mind-blowing but it's only been a month so its knowledge from there is also limited.

lat seems like it could be useful to cross-reference company-wide projects.

potentially. I'm toying with the idea of distributed lat ;)

I have a vitepress package in most of my repos. It is a knowledge graph that also just happens to produce heat looking docs for humans when served over http. Agents are very happy to read the raw .md.

managing agents.md is important, especially at scale. however I wonder how much of a measurable difference something like this makes? in theory, it's cool, but can you show me that it's actually performing better as compared to a large agents.md, nested agents.md, skills?

more general point being that we need to be methodical about the way we manage agent context. if lat.md shows a 10% broad improvement in agent perf in my repo, then I would certainly push for adoption. until then, vibes aren't enough

I'm working on a blog post and on benchmarks. Here [1] Armin suggested I take something like quickjs, built lat base for it, and compare side by side how, say, claude code works with lat vs. without.

I'm very early into this and need to build proper harness, but I can see sometimes lat allowing for up to 2x faster coding sessions. But the main benefit to me isn't speed, it's the fact that I can now review diffs faster and stay more engaged with the agent.

[1] https://x.com/mitsuhiko/status/2037649308086902989?s=20

Very cool, interested to read more once you post! FWIW I've been building eval infras that does something adjacent/related — replaying real repo work against different agent configs, and measuring the agent's quality dimensions (pass/fail, but also human intent alignment, code review, etc.). If you want to compare notes on the harness design, or if having an independent eval of lat vs. no-lat on quickjs would be useful, happy to chat :)

So... does it work? Good description of what it does, but, does it actually make agents better, or use less tokens? What's the benchmark?

I'm working on a blog post about lat, stay tuned. Yes, it does work and I continue improving it.

Cool. I've been thinking about tools to help coding agents lately, and I've always wondered if I could do a better job than actual Claude Code / Codex. They can do the `rg`s and kick off multiple agents and the skills, etc, and I couldn't find a way to actually prove there's a ROI of using something like Agent Lattice. Curious to see the results!

One anecdote: I added lat to quickjs codebase and asked questions like "how does quickjs gc circular object graphs" and was able to get an answer in 12 seconds vs 2 minutes without lat. But take this with a grain of salt, I need better benchmark harness. Benchmarks are hard to do in general, for LLMs it's its own hell. :)

i've had a related question recently but it didn't get much traction

https://news.ycombinator.com/item?id=47543324

what's the point of markdown? there's nothing useful you can do with it other than handing it over to llm and getting some probabilistic response

Read this section, I've just updated it to be more clear, hopefully will answer your question:

https://github.com/1st1/lat.md?tab=readme-ov-file#the-idea

> Test specs with enforcement — test cases can be described as sections in lat.md/ and marked with require-code-mention: true. Each spec then must be referenced by a // @lat: comment in test code. lat check flags any spec without a backlink, so you can review and maintain test coverage from the knowledge graph.

if you mean this paragraph - imo that's still too hand-wavy compared to enforcement through generative tests for spec conformance.

can you tell me more about what you mean by generative tests for spec conformance?

tests generated based on normative spec to validate if the implementation accepts the right inputs, rejects the wrong ones, produces correctly shaped outputs and to some extend also validates behavior.

normative specs allow generation of such tests deterministically - no need to spend tokens, no risk of hallucinations, much higher level of confidence that generated code is correct and much more accurate feedback from the test system into the LLM loop.

So we are reinventing the docs /*/*.md directory? /s I think this is a good idea just don’t really get why would you need a tool around it

One of the things that I've been chewing on lately is the sync problem. Having a CI job that identifies places where the docs have drifted from the implementation seems pretty valuable.

Python community figured this out in 2001: https://docs.python.org/3/library/doctest.html

I don't think this is related in any way.

> Having a CI job that identifies places where the docs have drifted from the implementation seems pretty valuable.

https://docs.python.org/3/library/doctest.html

> To check that a module’s docstrings are up-to-date by verifying that all interactive examples still work as documented. To perform regression testing by verifying that interactive examples from a test file or a test object work as expected. To write tutorial documentation for a package, liberally illustrated with input-output examples. Depending on whether the examples or the expository text are emphasized, this has the flavor of “literate testing” or “executable documentation”.

Seems pretty related to me.

Not really.

> Having a CI job that identifies places where the docs have drifted from the implementation seems pretty valuable.

Testing with lat isn't about ensuring consistency of code with public API documentation. It is about:

* ensuring you can quickly analyze what tests were added / changed by looking at the English description

* ensuring you spot when an agent randomly drops or alters an important functional/regression tests

The problem with coding agents is that they produce enormous diffs, and while reading tests code is very important in practice your focus and attention drifts and you can't do thorough analysis.

This isn't a new problem though, the same thing applies to classic code reviews -- rarely coding is a bottle neck, it's getting all reviews from humans to vet the change.

Lat shifts the focus from reading test code to understanding the semantics of the test. And because instead of reviewing 2000 lines of code you can focus on reviewing only 100 lines change in lat.md you'll be able to control your tests and implementation more tightly.

For projects where code quality isn't paramount I now just glance over the code to spot anti-pattern and models failing to DRY and resorting to duplicating large swaths of code.

I appreciate doctest very much, but those aren’t the kind of documentation I’m worried about drifting. I’m thinking more on the “how does this communication protocol between this server and this client work?”, which is generally terrible to try to summarize in doctests. If you want to take the idea to the extreme, imagine a CI test that answers “does this server implementation conform to this RFC?”

tl;dr: One file, bad (gets too big for context)

So give you agent a whole obsidian

I am skeptical how that helps. Agents cant just grep in one big file if reading entire file is the problem.

The problem is that for any non-trivial question agents have to grep A LOT to understand the high-level logic expressed in code. In big projects grepping can take minutes some times. Lat short-circuits grepping significantly with `lat search` that agents happily use.

Reply on news.ycombinator.com