Hacker News

Not knocking the article in any way but from the headline I was expecting - perhaps hoping - this would be about some innovation in filesystems research like it was the 90's again. That's not what this is.

It's about how filesystems as they are (and have been for decades) are proving to be powerful tools for LLMs/agents.

Does everyone just use AI to write these days? Or is the style so infectious that I just see it everywhere? I swear there needs to be some convention around labeling a post with how much AI was used in its creation.

I was having exact same observation, albeit from a bit diffrent perspective: SaaS. This is where as the code tends to be temporary and very domain specific, the data (files) must strive to be boring standards.

The problem today is that we build specific, short-lived apps that lock data into formats only they can read. If you don't use universal formats, your system is fragile. We can still open JPEGs from 1995 because the files don't depend on the software used to make them. Using obscure or proprietary formats is just technical debt that will eventually kill your project. File or forget.

Notable mention: Plan 9 from Bell Labs.

https://en.wikipedia.org/wiki/Plan_9_from_Bell_Labs

We once again discover that Plan9 and UNIX were right. The most powerful, lowest common denominator interface is text files exposed over a file system. Now to get back to making 9p2026.

The article gets some fundamentals completely wrong though: file systems are full graphs, not strict trees and are definitely not acyclic

I think this article just speaks to the immaturity of our use of AI at this "moment."

Production grade systems might be written by agents running on filesystem skills, but the production systems themselves will run on consistent and scalable data structures.

Meanwhile the UI of AI agents will almost certainly evolve away from desktop computers and toward audio/visual interfaces. An agent might get more context from a zoom call with you, once tone and body language can be used to increase the bandwidth between you.

Over a number of files similar to a codebase, that are well organized (like a codebase) the coding agents and harnesses are quite good at finding information, they clearly train on them so they will only improve.

The challenge is how to structure messy data as a filesystem the agent can use. That is a lot harder than querying a vector db for a semantic query.

The code bases we’ve been using agents in had been pruned and maintained over years, we’ve got principles like DRY that pushed us to put the answer in one place… implicitly building and maintaining that graph with all the actors in the system invested in maintaining this. This is not the case for messy data, so while I see the authors point and agree that a filesystem is a better structure for context over time, we haven’t supplanted search yet for non-code data.

Resonates deeply with me. I’ve moved personal data out of ~10 SaaS systems into a single directory structure in the last year. Agents pay a higher price for fragmentation than humans. A well-organized system of files eliminates that fragmentation. It’s enough for single player. I suspect we’ll see new databases emerge that enable low multi-player (safe writes etc) scenarios without making the filesystem data more opaque. Not unlike what QMD is for search.

In other words, file systems are an excellent way to organise information. I mean, yeah - we've been using them forever.

File systems are not a good abstraction mechanism for remote procedure calls, though. I think it's important to distinguish between the two, since I find there are a lot of articles conflating both - comparing MCPs to SKILLs, which are completely different things.

I think the confusion comes from the fact that MCP came before SKILLs, and there's a mental model where SKILLs are somehow "better than" MCPs. This is like saying local Word documents are better than a fully integrated collaborative office suite. It's just not the same thing.

The reason SKILLs work so well is because there's 50 years of accumulated knowledge of how to run rudimentary Unix tools.

the TLDR

File systems - organising information MCP/APIs - remote procedure calls

Digression: a file system is a terrible abstraction. The ceremonial file tree, where branches are directories and you have to hang your file on a particular branch like a Christmas ornament.

Relational is better. Hell, and kind of unique identifier would be nice. So many better ways to organize data stores.

I don't think there's a lot magical about files beyond (a) they are native for LLMs and coding because they both process text and (b)when things are rapidly in flux, unstructured formats prosper because flexibility is king. Literally any fixed format you try and describe becomes rapidly outdated and fails to serve the purpose. For example it feels like MCP is already ageing like milk.

Which is mainly to say, trust me, this is a temporary state, the god of complexity is coming. It is utterly inevitable. The people who created React, Kubernetes, all those Java frameworks you hated etc didn't go away. They are right now thinking about how amazing it would be if you if you stacked ten different tools together with brand new structured file formats and databases. We already have "beads" and "gastown" where this is starting. Enjoy these times because a couple of years from now it will already be the end of the "fun" part I think.

I wonder how much of a lost in the middle effect there is and if there could be or are tools that specifically differentiate optimizing post compaction "seeding". One problem I've run into with open spec is after a compaction, or kicking off a new session, it is easy to start out already ~50k tokens in and I assume somewhat more vulnerable to lost in the middle type effects before any actual coding may have taken place.

I thing the real impact behind the scenes here is Bash(). Filesystem relevance is a bit coincidental to placing an agent on an operating system and giving it full capability over it.

I'm not too deep into agentic coding, but I hadn't understood why people write `SOUL.md` files like no tomorrow. Does anyone think these will be called the same three years from now?

If you've got a coding convention, enforce it using a linter. Have the LLM write the rules and integrate it into the local build and CI tool.

Has noone ever thought about how – gasp – a future human collaborator would be onboarded?

Can we bring back Plan9 architecture now? It had what was essentially MCP. You make a custom device driver, and anything really can be a file. Not only that, but you network them, so a file on local disk could be a display on a remote host (or whatever). Just tell the agent to read/write files and it doesn't need to figure out either MCP or tool calls.

Funny, decades ago (mid-80s), I had to write a onetime fix on a what would be now a very low memory system, the data in question had a unique key of 8 7bit-ascii characters.

Instead of reading multi-meg data into memory to determine what to do, I used the file system and the program would store data related to the key in sub directories instead. The older people saw what I did and thought that was interesting. With development time factored in, doing it this way ended up being much faster and avoided memory issues that would have occurred.

So with AI, back to the old ways I guess :)

As TFA basically says: files on a filesystem is a DB. Just a very crude one. There aren't nice indexes for a variety of things. "Views" are not really there (arguably you can create different views with links but it's, once again, very crude). But it's definitely a DB, represented as a tree indeed as TFA mentions.

My life's data, including all the official stuff (bank statements, notary acts, statements made to the police [witness, etc.], insurance, property titels), all my coding projects, all the family pictures (not just the ones I took) and all the stuff I forgot, is in files, not in a dedicated DB. But these files are a definitely a database.

And because I don't want to deal with data corruption and even less want to deal with synching now corrupted data, many of my files contains, in their filename, a partial cryptographic checksum. E.g. "dsc239879879.jpg" becomes "dsc239789879-b3-6f338201b7.jpg" (meaning the Blake3 hash of that file has to begin with 6f338201b7 or the file is corrupted).

At any time, if I want to, I can import these in "real" dedicated DBs. For example I can pass my pictures as a read-only to "I'm Mich" (immich) and then query my pictures: "Find me all the pictures of Eliza" or "Find me all the pictures taken in 2016 on the french riviera".

But the real database of my all my life is and shall always be files on a filesystem.

With a "real" database, a backup can be as simple as a dump. With files backuping involve... Making sure you keep a proper version of all your files.

I'd say files are even more important than the filesystem: a backup on a BluRay disc or on an ext4-formatted SSD or on an exfat formatted SSD or on a tape... Doesn't matter: the files are the data.

A filesystem is the first "database" with these data: a crude one, with only simple queries. But a filesystem is definitely a database.

The main advantage of this very simple database is that as long as the data are accessible, you know your data is safe and can always use them to populate more advanced databases if needed.

Except android and iOS are both trying to keep you away from your own files.

Here’s me getting excited that a new file system is being developed but alas, just talk about text files.

Does this really have to do with file systems? Replacing RAG/context stuffing with tool calls for data access seems like the actual change. Whether the tool call is backed by a file system or DB or whatever shouldn’t matter, right?

nice, esp. liked - "our memories, our thoughts, our designs should outlive the software we used to create them"

It reminds me a lot of Hans Reiser’s original white paper, which can be found at https://web.archive.org/web/20070927003401/http://www.namesy.... Add some embeddings and boom.

This article said some things I couldn’t put into words about different AI tools. Thanks for sharing.

I revived my Johnny Decimal system as my single source of truth for almost anything and couldn't be happier. The filing is done mostly by agents now but I still have the overview myself.

Great read. Thanks for sharing

I don't think this paradigm will last, or be what becomes the more common structure in the future. This will still suffers from conflicts of persona and objective, plus has the issue that individual apps will need protected file hierarchies to prevent malicious injections. I don't see this as a solution, just a deck chair shuffle.

I've been researching and building with a different paradigm, an inversion of the tool calling concept that creates contextual agents of limited scope, but pipelines of them, with the user in triplicate control of agent as author, operator of an application with a clear goal, and conversationally cooperating on a task with one or more agents.

I create agents that are inside open source software, making that application "intelligent", and the user has control to make the agent an expert in the type of work that human uses that software. Imagine a word processor that when used by a documentation author has multiple documentation agents that co-work with the author. While that same word processor when used by a, for example, romance novelist has similar agents but experts in a different literary / document goal. Then do this with spreadsheets, and project management software, and you get an intelligent office suite with amazing levels of user assistance.

In this structure, context/task specific knowledge is placed inside other software, providing complex processes to the user they can conversationally request and compose on the fly, use and save as a new agent for repeated use, or discard as something built for the moment. The agents are inside other software, with full knowledge of that application in addition to task knowledge related to why the user is using that software. It's a unified agent creation and use and chain-of-thought live editing environment, in context with what one is doing in other software.

I wrap the entire structure into a permission hierarchy that mirrors departments, projects, and project staff, creating an application suite structure more secure than this Filesystems approach, with substantially more user controls that do not expose the potential for malicious application. The agents are each for a specific purpose, which limits their reach and potential for damage. Being purpose built, the users (who are task focused, not developers) easily edit and enhance the agents they use because that is the job/career they already know and continue to do, just with agent help.