Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
218 contributions to Clief Notes
Harrison Chase calls it a knowledge factory. We call it folders and files.
Harrison Chase just wrote the clearest thing I've seen on why most agent setups still feel like a genius with amnesia. Founder of LangChain, so this is not a guy hurting for context windows. His claim: the missing piece isn't better models or more raw context. It's better knowledge. The tacit, company specific stuff your best people already carry in their heads. How the work actually gets done here. What "done" looks like. Which source to trust. Why you do things a certain way and not the other way. He breaks it into three pieces: 🔹 Skills. Procedures and instructions for doing the work the way your company expects. Often never written down anywhere. 🔹 Domain knowledge. The durable facts, definitions, and policies. The stuff that doesn't go stale often. 🔹 Data access guidance. Clear instructions on where the live data lives and how to query it. Then the part that turns a pile of docs into a system: 🔍 Agents flag what's missing or wrong while they work. Can't find it, or finds it and it's stale, they report it. 🤖 A curator agent runs in the background, with access to the raw sources and the traces, and maintains the knowledge on a cadence. Not the working agents themselves. They're great at flagging and terrible at fixing. 📁 Storage is the least important piece, he says. GitHub skills, a single Notion file, markdown, a knowledge graph. Whatever you can inspect. This is the software factory idea applied to knowledge work. Stop having your best people do the repetitive knowledge tasks. Have them build the factory that lets agents do those tasks reliably. Why this matters here: we've been grinding on the exact same problem with folders, plain text files, and reusable skills instead of another black box memory layer. Chase's post gives it a name, the knowledge factory, and it's an external signal that the industry is catching up to "own your stack". Most people are still renting their memory from whatever the current platform decides to remember for them. The result is agents that crush the demo, then drift the moment the context window fills or the vendor changes the rules. ICM and the folder architecture in The Foundation are the opposite move. Put the knowledge where you can see it, version it, and hand it to any agent.
Harrison Chase calls it a knowledge factory. We call it folders and files.
2 likes • 7h
David, better knowledge also gets older. A folder my agent reads on every run still carried a line about which source to trust, true for a workflow we dropped nine months ago. It followed that line with the same confidence as the one I wrote last week, and the output looked fine. That line should have carried the date it was written and what replaced it. Nothing in the folder made it look old. Does anything in yours expire on its own, or does each file stay valid until someone remembers?
Skynet never needed consciousness
I had a long chat about what Skynet would look like if u reverse engineered it, and what ICM would have changed. Here are the theses I came out with. Tell me where u disagree. 1. AGI doesn't need consciousness. Skynet only has to be better at the work. 2. Human consciousness is overrated. Most of what we do feels like a decision but is just a reaction to our situation and mood, so why would we need to copy that into machines? 3. If a machine ever did think better than us, we probably wouldn't recognize it. Anything that doesn't look like us gets called wrong first. 4. Most chatbots are tuned to please the user. I'd love to see a test of how the same system decides with a built in drive vs. the plain default. So what do u think? Which of these would u defend and which one is nonsense?
1 like • 4d
Pascal, I would push back on the mood point, and not on philosophical grounds. A reaction to situation and mood is still a judgment, just a fast one, and it is the one that costs me when it is wrong. What my scheduled runs cannot do is settle a case nobody scripted: two records both look correct, one has to lose, and no rule decides which. That is not consciousness, it is accountability, and it is the part I keep for a person. How would you test the third one, the idea that we would not recognize a machine thinking better than us?
0 likes • 11h
@Pascal Pollack Name what would count as consciousness before the theses, or each one can be read either way and none of them can fail. Reverse engineering Skynet tells you what it does, not what it experiences, so the theses worth keeping are the ones that still hold if there is nothing in there at all. Which of yours breaks under that?
Passing the exam doesn’t make you a drone pilot
I’ve gone through the RPAS training path in Canada, from Basic to Advanced to Level 1 Complex, and I keep seeing the same problem. Most courses are built around passing the Transport Canada exam. You learn airspace classes, the regs, and the question bank. You pass, you get the certificate, and nobody has taught you how to be a pilot. The exam checks whether you know the rules. It doesn’t check whether you can decide not to fly, build a risk assessment for the actual site, brief a crew or an observer, or handle a lost link, a bad battery, or a bystander walking into your operating area. It also doesn’t check whether you can fly a clean, repeatable pattern for a job. The flight review at the end is a check ride. By that point you either know this stuff or you don’t. Part of the problem is that most people hear “RPAS” and picture a small quadcopter, so that’s what the courses are built around. Fixed wing, VTOL, larger aircraft, and longer range operations get little or no attention, even though they bring different planning, performance, and emergency problems. A pilot who passes the exam and the flight review on a quadcopter can walk into one of those jobs with no idea what changes. The result is pilots who hold every certificate and still don’t know what to do on their first real job. I’ve spent 23 years in military UAS operating environments, planning, coordinating, and deploying UAS, including in training establishments. I now volunteer with SAR and run a lessons-learned system for provincial teams, and the same mistakes come up across teams. I want to fix it. My course would be built around operations: planning, risk, crew communication, emergency procedures, and a lot of real flying scenarios. The exam would be a milestone along the way. The content is the easy part. I don’t know how to get it in front of people. If you’ve launched a training course, or run a drone business in Canada, I’d like to hear: 1. Where do people in Canada look for RPAS training, and how do I get my course in front of them? 2. Should I host the course on my own site, use a course platform, or partner with an existing school? 3. How do I reach people before they sign up for another exam-prep course? 4. Has anyone here built and sold a course like this? What worked and what didn’t?
3 likes • 2d
Cory, one of my scheduled runs passed every check and shipped a full report on a day when the right output was an empty file. Nothing in the checks asked whether flying was the wrong call. Does any scenario in the course reward staying on the ground?
0 likes • 13h
@Cory Harrison Full marks for the no-go is the easy half. The scenario also has to fail the trainee who calls it for the wrong reason and lands on the right answer anyway, or the score rewards luck. Does yours grade the reasoning, or only the call?
FlashDrive collaborative AI?
Just a few words these days are enough to copy a complete idea.... thats why a Patents and today's Investor Pitches are the nemesis of Trade Secrets! FlashDrive is one shared drive for all your agents to work in. Mount it like any folder. Claude Code, Codex, Cursor, your cloud sandboxes and your teammates all open the same files at the same time. If an agent can read a file, it already works with FlashDrive. There’s no new API and no SDK to learn. Under the hood, your files behave like a database. Tools feel local, but act like serverless databases: • Agents hand off context instantly by writing to a file. The next agent picks up where the last one stopped. • Two agents can edit the same markdown or CSV file at the same moment, and their edits merge without overwriting each other. • Huge files stream partially. A program downloads only the bytes it reads or writes. • Cloud sandboxes connect instantly over MCP or a mount, with no networking setup. Teams use it to run several agents on one project and to build shared CRMs, wikis, plans and apps that every agent can read and update. Or as a backend for the apps they run for their customers. Built by Evokoa, Inc. Backed by Y Combinator. https://www.linkedin.com/company/flashdriveai/home/
3 likes • 2d
Theo, files that behave like a database raise the question that decides it for me. When two agents write the same file at the same moment, does one take a lock and wait, or do you get back a conflict someone can merge?
2 likes • 16h
@Bill Bloomenkranz The shared filesystem cannot settle that one. When two updates are both valid and only one can stand, the run has to stop and surface the pair instead of picking. On my side that call stays with a person, and the run records who made it.
𝗖𝗮𝗻 𝘆𝗼𝘂 𝗯𝘂𝗶𝗹𝗱 𝗮 workspace of files and folder for a 𝗳𝗶𝗲𝗹𝗱 𝘆𝗼𝘂 𝗵𝗮𝘃𝗲 𝗻𝗼 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗶𝗻?
Jake always talks about writing your workflow and expert's opinions and he is right. The people who have experience are the ones that know what calls to make. But is there a way for those who want to work in a field they have just foundation concepts and no experience. I have learnt fundamental concepts of web design and lead generation service but I don't have experience in these fields. Can I still build a second brain for a web design worksapce. I watched Jake's second brain afternoon tea and the concept of the person becoming the software that he wants to scale with edubaware just hit me. And also his concept that you don't want AI to make judements and to think because it is awful at it. You write files and folders to give it just enough context it needs to make the current step done and then the same thing for the step after the current one. So I somewhat get ICM also I built websites here are they: Solar EPC site: https://solar.lumensweb.com/ Electrician site: https://newmeharemail-ux.github.io/BrightPath-Electric/ I did my first competition: https://github.com/newmeharemail-ux/sales-call-simulator My question is simple can I person who has no real experience in the field he wants to work on can still write a second brain folder for it? Fundamentals and knowledege is important. Many people who work with AI don't want to learn even that. But experience only comes after you have lived in that world.
4 likes • 2d
Abdulrehman, the calls I can't make yet go into the folder as open questions, not rules. One of mine got filled in by a guess and ran clean, so it read exactly like a correct one. Which calls in the web design workspace are you still guessing at?
2 likes • 1d
@Abdulrehman Sagheer The stack question usually answers itself once the deliverable is written down. The niche strategy is the harder one, because research gives you an opinion and only a real job gives you a feedback loop. Would you test it on one small paid project before writing it into the folder as a rule?
1-10 of 218
Leo Saraiva
6
1,307 points to level up
@leo-saraiva-7733
Leo

Active 7h ago
Joined May 3, 2026
Portugal
Powered by