I don't understand the point of automating note taking. It never worked for me to copy paste text into my notes and now you can 100x that?
The whole point of taking notes for me is to read a source critically, fit it in my mental model, and then document that. Then sometimes I look it up for the details. But for me the shaping of the mental model is what counts
First of all, this is more than just note taking. It appears to be a (yet another) harness for coordinating work between agents with minimal human intervention. And as such, shouldn’t part of the point be to not have to build that mental model yourself, but rather offload it to the shared LLM “brain”?
Highly debatable whether it’s possible to create anything truly valuable (valuable for the owner of the product that is) with this approach, though. I’m not convinced that it will ever be possible to create valuable products from just a prompt and an agent harness. At that point, the product itself can be (re)created by anyone, product development has been commodified, and the only thing of value is tokens.
My hypothesis is that “do things that don’t scale”[0] will still apply well into the future, but the “things that don’t scale” will change.
All that said, I’ve finally started using Obsidian after setting up some skills for note taking, researching, linking, splitting, and restructuring the knowledge base. I’ve never been able to spend time on keeping it structured, but I now have a digital secretary that can do all of the work I’m too lazy to do. I can just jot down random thoughts and ideas, and the agent helps me structure it, ask follow-up questions, relate it to other ongoing work, and so on. I’m still putting in the work of reading sources and building a mental model, but I’m also getting high-quality notes almost for free.
The few scientific studies out there actually show a degradation of output quality when these markdown collections are fully LLM maintained (opposed to an increase when they’re human maintained), which I found fascinating.
I think the sweet spot is human curation of these documents, but unsupervised management is never the answer, especially if you don’t consciously think about debt / drift in these.
Are you referring to the one (1) study that showed that when cheaper LLM's auto-generated an AGENTS.md, it performed more poorly than human editted AGENTS.md? https://arxiv.org/abs/2602.11988
I'd love to see other sources that seek to academically understand how LLM's use context, specifically ones using modern frontier models.
My takeaway from these CLAUDE.md/AGENTS.md efforts isn't that agents can't maintain any form of context at all, rather, that bloated CLAUDE.md files filled with data that agents can gather on the spot very quickly are counter-productive.
For information which cannot be gathered on the spot quickly, clearly (to me) context helps improve quality, and in my experience, having AI summarize some key information in a thread and write to a file, and organize that, has been helpful and useful.
i've been running a variation of the _llm writes a wiki_ since late february. i run it on a sprite (sprites.dev from fly.io), it's public but i don't particularly advertise it. i completely vibe coded the shit out of it with claude. the app side and the content. the app side makes the content accessible to other agent instances, lists some documents at the root, provides search function, and let's me read it on a browser with nice typography if i want to, as opposed to raw markdown.
it's neat, i can create a new sprite/whatever, point claude at the root, and tell it to setup zswap and it will know exactly how to do so in that environment. if something changes, and there's some fiddling to make it work, i can ask it to write a report and send it in to fold into the existing docs.
How do you anticipate teams deploying this? I’m wary of GitHub for sensitive business documents, and wonder what an easy secure agent friendly deployment looks like. Cloudflare or GCP are maybe good candidates
Put AI in your product name, make billion dollars. Put Karpathy in your blog article, get hired by Anthropic as Principal engineer. Milk money as long as fad last. No one is thinking about customer needs, everyone is trying to wash hands in the wave as it last.
Just like NFTs, just like the blockchain before that, in some ways kind of like the Web 2.0 craze (though we at least built some things then and the tight financing at the time kept a lid on it).
This LLM stuff at least has some real possibilities and value, and is very fun tech to learn about and play with.
I long ago accepted that there’s money to be made, as long as it’s not unethical, then get involved. Can build cool things that do have value, while enjoying the VC/PE money sloshing around
The space of self building artefacts is interesting and is booming now because recent LLM versions are becoming good at it fast (in particular if they are of the "coding" kind).
I've also experimented recently with such a project [0] with minimal dependencies and with some emphasis on staying local and in control of the agent.
It's building and organising its own sqlite database to fulfil a long running task given in a prompt while having access to a local wikipedia copy for source data.
A very minimal set of harness and tools to experiment with agent drift.
Adding image processing tool in this framework is also easy (by encoding them as base64 (details can be vibecoded by local LLMs) and passing them to llama.cpp ).
It's a useful versatile tool to have.
For example, I used to have some scripts which processed invoices and receipts in some folders, extracting amount date and vendor from them using amazon textract, then I have a ui to manually check the numbers and put the result in some csv for the accountant every year. Now I can replace the amazon textract requests by a llama.cpp model call with the appropriate prompt while still my existing invoices tools, but now with a prompt I can do a lot more creative accounting.
I have also experimented with some vibecoded variation of this code to drive a physical robot from a sequence of camera images and while it does move and reach the target in the simple cases (even though the LLM I use was never explicitly train to drive a robot), it is too slow (10s to choose the next action) for practical use. (The current no deep-learning controller I use for this robot does the vision processing loop at 20hz).
LLM models and the agents that use them are probabilistic, not deterministic. They accomplish something a percentage of the time, never every time.
That means the longer an agent runs on a task, the more likely it will fail the task. Running agents like this will always fail and burn a ton of token cash in the process.
One thing that LLM agents are good at is writing their own instructions. The trick is to limit the time and thinking steps in a thinking model then evaluate, update, and run again. A good metaphor is that agents trip. Don't let them run long enough to trip. It is better to let them run twice for 5 minutes than once for 10 minutes.
Give it a few weeks and self-referencing agents are going to be at the top of everybody's twitter feed.
The BM25-first routing bet is interesting. You mention 85% recall@20 on 500 artifacts, but the heuristic classifier routing "short lookups to BM25 and narrative queries to cited-answer" raises a practical question: what does the classifier key on to decide a query is narrative vs short? Token count? Syntactic structure? The reason I ask is that in agent-generated queries, the boundary is often blurry - an agent doing a dependency lookup might issue a surprisingly long, well-formed sentence. If the classifier routes those to the more expensive cited-answer loop it could negate the latency advantage of BM25 being first.
Re classifier routing: text-shape signals (token count, syntactic markers) underspecify the boundary, especially for agent-generated queries. The signal that worked better in our policy-gated tool-call setting was the surrounding intent context the agent was operating under, not the query string itself. An agent in a "fact-check" context emits long, well-formed sentences that actually want exact-match retrieval; an agent in an "open research" context emits surprisingly short queries that need narrative retrieval. If the runtime can read the tool or skill context at query time, routing on that is less ambiguous than text shape. Doesn't help if the wiki is a black-box MCP server with no caller-side context, but it's worth offering an optional context hint in the lookup payload.
Any particular reason for BM25? Why not just a table of contents or index structure (json, md, whatever) that is updated automatically and fed in context at query time? I know bag of words is great for speed but even at 1000s of documents, the index can be quite cheap and will maximise precision
I read the durability thing as markdown files are very open, easy to find software for, simple and are widely used. All of this together almost guarantees that they will he viewable/usable in the far future.
We've [nex, creators] have been using it as a sounding board. I think that in its current state it's actually more useful for someone to learn about how to run a business - "what does a CEO vs PM do" and/or learn about the pros/cons of running a bunch of agents at once.
The whole point of taking notes for me is to read a source critically, fit it in my mental model, and then document that. Then sometimes I look it up for the details. But for me the shaping of the mental model is what counts
Highly debatable whether it’s possible to create anything truly valuable (valuable for the owner of the product that is) with this approach, though. I’m not convinced that it will ever be possible to create valuable products from just a prompt and an agent harness. At that point, the product itself can be (re)created by anyone, product development has been commodified, and the only thing of value is tokens.
My hypothesis is that “do things that don’t scale”[0] will still apply well into the future, but the “things that don’t scale” will change.
All that said, I’ve finally started using Obsidian after setting up some skills for note taking, researching, linking, splitting, and restructuring the knowledge base. I’ve never been able to spend time on keeping it structured, but I now have a digital secretary that can do all of the work I’m too lazy to do. I can just jot down random thoughts and ideas, and the agent helps me structure it, ask follow-up questions, relate it to other ongoing work, and so on. I’m still putting in the work of reading sources and building a mental model, but I’m also getting high-quality notes almost for free.
[0]: https://www.paulgraham.com/ds.html
I think the sweet spot is human curation of these documents, but unsupervised management is never the answer, especially if you don’t consciously think about debt / drift in these.
I'd love to see other sources that seek to academically understand how LLM's use context, specifically ones using modern frontier models.
My takeaway from these CLAUDE.md/AGENTS.md efforts isn't that agents can't maintain any form of context at all, rather, that bloated CLAUDE.md files filled with data that agents can gather on the spot very quickly are counter-productive.
For information which cannot be gathered on the spot quickly, clearly (to me) context helps improve quality, and in my experience, having AI summarize some key information in a thread and write to a file, and organize that, has been helpful and useful.
it's neat, i can create a new sprite/whatever, point claude at the root, and tell it to setup zswap and it will know exactly how to do so in that environment. if something changes, and there's some fiddling to make it work, i can ask it to write a report and send it in to fold into the existing docs.
This LLM stuff at least has some real possibilities and value, and is very fun tech to learn about and play with.
I long ago accepted that there’s money to be made, as long as it’s not unethical, then get involved. Can build cool things that do have value, while enjoying the VC/PE money sloshing around
I've also experimented recently with such a project [0] with minimal dependencies and with some emphasis on staying local and in control of the agent.
It's building and organising its own sqlite database to fulfil a long running task given in a prompt while having access to a local wikipedia copy for source data.
A very minimal set of harness and tools to experiment with agent drift.
Adding image processing tool in this framework is also easy (by encoding them as base64 (details can be vibecoded by local LLMs) and passing them to llama.cpp ).
It's a useful versatile tool to have.
For example, I used to have some scripts which processed invoices and receipts in some folders, extracting amount date and vendor from them using amazon textract, then I have a ui to manually check the numbers and put the result in some csv for the accountant every year. Now I can replace the amazon textract requests by a llama.cpp model call with the appropriate prompt while still my existing invoices tools, but now with a prompt I can do a lot more creative accounting.
I have also experimented with some vibecoded variation of this code to drive a physical robot from a sequence of camera images and while it does move and reach the target in the simple cases (even though the LLM I use was never explicitly train to drive a robot), it is too slow (10s to choose the next action) for practical use. (The current no deep-learning controller I use for this robot does the vision processing loop at 20hz).
[0]https://github.com/GistNoesis/Shoggoth.db/
That means the longer an agent runs on a task, the more likely it will fail the task. Running agents like this will always fail and burn a ton of token cash in the process.
One thing that LLM agents are good at is writing their own instructions. The trick is to limit the time and thinking steps in a thinking model then evaluate, update, and run again. A good metaphor is that agents trip. Don't let them run long enough to trip. It is better to let them run twice for 5 minutes than once for 10 minutes.
Give it a few weeks and self-referencing agents are going to be at the top of everybody's twitter feed.
https://x.com/karpathy/status/2039805659525644595
https://xcancel.com/karpathy/status/2039805659525644595
But also would like to understand how markdown helps in durability - if I understand correctly markdown has a edge over other formats for LLMs.
Also I too am building something similar on markdown which versions with git but for a completely different use case : https://voiden.md/
1 -https://x.com/__endif/status/2039810651120705569
Every time I hear someone say "I have a team of agents", what I hear is "I'm shipping heaps of AI slop".