This Guy Writes Books For Agents (CandleKeep, Sahar Carmel)

Añadido: 2026-05-09

[00:00:00]Hello, Sal. Hey, Tom. How are you? Good, good. Yourself? I'm fine, thank you.

[00:00:05]Thank you so much for coming on the show. For those of you who don't know, Sal is the founder of Kindle Keep, which is a product I reviewed on the channel, I think by the time this goes out a few days or a week ago. Uh Sal came to say two things. One, I was incorrect in my video in some portions, so he will kind of illustrate where I was wrong. And two, to kind of give you the story behind it, because I kind of dove into the tool, but not into like the reason behind it it even existing to begin with. Uh so, I'm going to shut up, but I'm going to just just kind of Sal, if you can present yourself, tell us a little bit about who you are, why you built Kindle Keep, like why is this whole thing happening, uh and I'll intervene with annoying questions in a bit. So. All right. So, hi everyone. I'm Sal Carmel. I'm 34 years old. I'm the director of AI enablement at a company called Mix Tiles. I have a community revolving around agents and AI usage, um in general. And I also developed Kindle Keep as my um tool for AI agents, actually. So, um back in the days, about 5 months ago, I've been building with Cloud Code in four terminals in my uh in my PC, and I find myself having time to maybe play something. And I'm a video fan of Dungeons and Dragons, and I was thinking how I can play Dungeons and Dragons in my terminal, in my Cloud Code, right? Which is a totally logical thing for a random person [laughter] to say. Like, what I want to do? I'm going to play D&D in my Sure, of Sorry, go ahead. Yeah, no. Yeah.

[00:01:35]Uh so, um what I was thinking that Dungeons and Dragons is a books book-based game, right? You have the adventure, you have the um um Dungeon Master Guide, and you have the Player's Handbook. Um and you can ask Claude to generate some kind of adventure, or try to estimate what was in the adventure book, original book.

[00:01:56]But, you will find yourself with a bunch of mediocre or some vague memory of Claude, of what he was reading about Dungeons and Dragons, and specifically the adventure. And I was thinking how I can create tools that will let my agents read books, and also uh as a side effect, will let me play Dungeons and Dragons while I code with Claude code. So, I've created a bunch of tools that will help agents to properly and easily read books.

[00:02:27]Uh I've um I've done three kind of tool tools. One is get library, one is get table of content, and one get pages, right?

[00:02:38]>> Maybe just to to stop you and to clarify for those of you who might have not watched the video, Kindle Kit comes with a web app and a CLI and like an um I don't think it's an MCP server. It's an Are those MCP tools? No, no. It's a CLI.

[00:02:49]>> CLI, right. So, your agent can call the CLI and actually get the library, get the book, get the chapter, whatever.

[00:02:56]It's essentially a library for your agent. Like, this is the best way to to to call it. Um >> Exactly. And if you >> as a library. There you But, now it's it's kind of a Kindle. I'll get to it.

[00:03:07]Yeah, please. Go ahead. Anyway, so imagine that you are a marketing guy and you have all your knowledge of how to do marketing in your company, right? If you would ask Claude for tips about how to do the go-to-market to your company, it would give you some kind of a mediocre or some kind of average answer from all what Claude has been trained on. But, if you give the agent your marketing guidelines, your playbook, whatever, you know, all the material that have been accumulated in your company, Claude will give you will give you much more curated answers based on your context, right?

[00:03:46]So, Kindle Kit started as a way to for me to play Dungeons and Dragons, but now teams have been using it to upload context to it, a company context, and make it available to all the agents in the company. You know what this is?

[00:04:01]Sorry for stopping you, but you know what this reminds me every company that I've ever worked in, especially startup companies, have this employee handbook thing. Now, nobody reads the employee handbook. Like nobody ever opened it.

[00:04:10]Imagine that we have like actual handbooks of each department where you can have a book, an actual book of what's happening inside the department.

[00:04:18]This is like we've been trying to do it for years, but nobody has time to do it all of these material handbooks. So, now we do, in a sense.

[00:04:26]Uh funny that you say that, but actually reading is cool, right? You can instantly give your agent knowledge that you don't possess. For example, if I'm a marketing guy and I want to do a cybersecurity uh audit to my Vibe Coded app, I can go and get a cybersecurity book, and instantly my agent knows something that I don't know.

[00:04:49]And I've implemented into Candlekeep a micro learning micro learning experience where the agent knows what he needs to do, and you the user can get some snippets of what they've been done, right?

[00:05:00]>> Okay. Um and now I've also added something to Candlekeep that aside from just uploading documents, you can also assign a thing you want to write about. For example, how I use Cloud Code for go-to-market. This is the title of what I want to to write. And then every time I use Cloud Code or my co-work to generate some work on go-to-market using AI, my agent will come up and say to me, "Hey, this looks like a nice thing to write about. You want me to add it to your book?"

[00:05:35]So, what you're saying essentially is you created the I think I I saw the post. It's about like the Andre came Andre Kapathy came out with the whole LLM wiki thing, and he's doing this with Obsidian and the agent. And you're saying that you can self-record or self-journal in a sense of all the process that you you're doing with agents by essentially creating a book of the way you do things. Kind of like a And the reason why I love this, you also called it manuscripts, which is just the best name for this. This is great marketing there, dude. Um I I and I think I think this is like a good time to ask because I really love the idea, but it it makes a lot of technical assumptions. So, walk with us a little bit about why the current way context is being built for agents. Rag is a very popular one, just dropping markdown files into a folder, create like um using something like context seven to get a docs uh references from whatever. Like, walk me through why books is the right abstraction for knowledge for agents in your opinion. Like, the technical specificity, if you don't mind.

[00:06:35]>> Yeah. So, if you want to keep in theme, uh I love books and I've read Antifragile, right?

[00:06:41]>> Yes. And Antifragile um Good book.

[00:06:44]>> is talking about the time um the time test, right? If something has survived enough time, like religion, has uh survived like 2,000 years, it is suspected to survive for another 2,000 years, right? So, if we're trying to assess what will be the best way to represent knowledge, Mhm. books have been with us for as long as the humanity can remember or actually from where history have begun. So, it have been evolved to be the best way to preserve context and now to access it, right? Because if you have a long book and you want to find some kind of occurrences in the book, you have the index at the end, right? And instantly you can go and see the sign of uh definition all throughout the book. Mhm.

[00:07:36]And table of content is a good way to do a procedural or uh progressive disclosure like agents want to do, right? If you have a skill, you want to do a progressive disclosure. So, ultimately, library is the ultimate way of progressive disclosure when I'm researching.

[00:07:55]Uh another thing is uh I'll stop you there. I'll stop you there. I'm sorry for the other thing. Um when we talk about abstraction in the world of programming, we usually try and mimic or make analogy between something we're building an architecture decision that we make to some pattern in the real world. Um database indexes come to mind in this in this way because what do you do when you like an index helps you search through a database essentially.

[00:08:18]So, what you're saying is and that is a an hypothesis, like it's the thing you built Candlekeep around. You're saying instead of letting it, for example, look at the tree file at the file tree in your in a drive and then read the front matter of markdown files or, for example, use something like um I'm lacking on the name. There's another thing that's like an SQLite-based database of like contacts that you can put into it. Or, you know, talk to something like Pine Tree, which is like a vector database, and query that again.

[00:08:44]What you're saying is for the agentic time that we are in at the moment, a good way to abstract away the problem of knowledge searching and knowledge indexing is exactly how we do it in human form. Table of contents.

[00:08:56]>> Exactly. Do you have some metrics on like whether it's actually better than the original rug stuff? Do you know any numbers?

[00:09:04]Um I've never done one, but I get a lot of feedbacks from people there's some user in Candlekeep that's been doing research on biblical resources, right? And if you ask your Claude if you would take all the Gemara, the Bible, the whatever, and put it in a rag, there's no way that you can actually ask something like what was the I don't know the Rambam asked about whatever, right? And you would to it right because semantic is not the only way you can assign stuff. It's semantic it's not the right way we as humans when I tell you something you're not you're not just instantly go and try to find something similar verbally.

[00:09:52]You have associations. You have some kind of a map in your knowledge about which books is relevant to what, right?

[00:10:00]So I haven't done benchmarks uh but for example when I try to um do a security review of a vibe coded app I've seen that I get 10x more vulnerabilities found against Claude's security review.

[00:10:16]Yeah. Why?

[00:10:18]It's it's very easy to understand because security is just a bunch of rules that I want to rule out, right?

[00:10:25]And Claude trying to do that from prompting and try to you know try to assess what he was learning. But if I give him a full book 600 rules of security when he have that in his possession he will do a a much better job. So benchmarks are not yet part of the of the claim. Honestly it just feels right. Uh you if you want to see something that's very similar cursor have started with rugs and now they they reverted into agentic search like Claude Codo.

[00:10:58]Right? So this is ridiculous. Again, I'm sorry for stopping you. I'm just very excited about this. This is like the the you said a multiple different things that are very interesting. So I'm just going to focus on the thing that really caught my attention. You mentioned that semantic is not the only way to search.

[00:11:11]Now we have been accustomed to the fact that when you search semantic search is the holy grail. It's expensive. It brings in great results. It checks all the boxes and we have to degrade to non-semantic search when we don't want to use you know the overly complex processes to create the ability for a semantic layer for a piece of content.

[00:11:30]What you're saying is wait semantic is not the whole the whole story. Sure, you want an understanding of the content not to allow something to look through it, but the correlation and what leads to what and whether this is newer and this is older and this comes from this side of the discipline, this comes from another school. This is a whole different layer of understanding that you can't get until you do something like reading a book because um you gave I love I love the the there's like um I don't know who I talked to maybe uh Oh, right. Matan. Matan who will be on the channel as well very soon. Uh who's like he's a fraud guy. He works at PayPal and he's like deep into the AI weeds and he was like um he has like a massive library on Kindle kit like 60 books or something.

[00:12:13]>> Massive.

[00:12:14]200. 200, sorry. It's like very big. And I asked him like what like how do you know which book to choose? And he said something that I loved. And I also have a library on the back here so I very much relate to the whole story.

[00:12:25]Sometimes you pull an Alex Hormozi book when you want to learn about offers, but the little of course I can't find it right now, but I used [laughter] to have it here somewhere. Uh the little red book of selling, which is a very well known little manuscript that is like from I know the '70s or something. And it talks about a time where you did sales over the phone in a specific way and it has concepts that people who called call today have already forgotten. It's not listed in any newer book, but it's the right thing to say at the right moment in time. And a semantic layer would not find it. It's something that you know because you know type thing. I I Exactly. No, sorry. Uh certainly. So, uh by the way, semantic layers and rags have came to our knowledge when agents and LLMs couldn't use tools. True.

[00:13:15]True. So, now that agents can use tools, there are much powerful and easier ways to get context into agents. For example, the most easy stuff is like um back in the days I was dealing with text to SQL.

[00:13:30]Now it's pretty much solved. I just give my agent an access to the database and it can just iterate and do whatever SQL it thinks about and just compiles it. So actually, my thought is that rags are a thing of the past and now the tools are being developed and LLMs are much better at using them. Actually, using tools is a better way to bring context into agents. And when you try to build a tool, then go ahead and think how humans use this tool and then mimic that. I'm calling that a human-driven development, actually.

[00:14:07]>> So interesting. So interesting and I want to I want to kind of take us because you're clearly a forward-thinking person. Like you're looking at this from the perspective of someone that is in the weeds, working on things.

[00:14:20]You have an amazing Nanoclu book. You have a bunch of different things that you do. And um I think what I want to take like for the few minutes that we have left, I want to take it to the way the world is going to be in your opinion. So I'll ask a very concrete question. Um you deal with context and knowledge acquisition for agents in a sense.

[00:14:40]We saw kind of looping back through latest things that have happened. We saw GPT-2 GPT-2 releasing with like the first real interesting LLM instruct model implementation that we saw in the wild. Then later on when we caught the Anthropic harnesses applications, yada yada, the cloud codes of the world, we saw another leap of oh, this is a tool and a harness that we can use in a different way than chat. Now we have MCP, A2A, a bunch of different protocols which allow these agents to communicate with other things. You are saying context is a big problem which many different people are trying to solve now. Something some standout will rise above the others like, you know, in all in all of these situations and we'll figure out what to do. Um, what's the next big problem? Like, what do you think is the next big thing that people are going to hit the wall on when they're trying to scale these things?

[00:15:33]Well, now that I've been building agents in production for about a year or so, um, I think that the next software will be building agents.

[00:15:43]I think building agents will not be solved. It will be just part of what we know of as software. You will deploy agent, it will do stuff. Sometimes it will fail. You will debug that and so on and so forth, right? So, agents essentially are trying to mimic some kind of uh minor abilities of a human.

[00:16:04]And in my uh in context is all have been always been the problem, right?

[00:16:11]Um, right now we're accumulating new data every day.

[00:16:15]Agents don't know how to build agents, right? And agents don't know about memory and they have even a little knowledge about MCPs. So, I think that the the next phase uh as humans written books for humans, the next big thing will be humans writing books for agents and agents writing books for agents because books is the most efficient way to deliver knowledge, right? So, if I'm trying to build the most prominent go-to-market agent, I'm a software developer, right?

[00:16:53]Yeah. And I'm trying to build that. Tom Gerhardt, for example, can take Kindle Keep and start writing the ultimate guide for agents how to be a go-to-market agent.

[00:17:04]And it's a living book, right? And instantly you can teach my agent how to do go-to-market.

[00:17:12]How to do go-to-market for agents.

[00:17:15]Agents can accumulate knowledge about what e have been trying to do as an agent to try and get customers and they will build upon that. So, I think that the future is also about context. I think that building agents will be the next thing. This will be the next software. And people will work around the clock making agents better. And I think you don't have many right now you don't have a a huge um arsenal of things to do with agents. You have sub agents, plugins, context, and that's it.

[00:17:52]Right? Everything is about context. But this is like the tools we have. And if I can take your agent and skyrocket it to have a better knowledge from what my agents have learned, Yeah. I can instantly make your agent better. 100%.

[00:18:09]Uh so I'm I'm going to stop you there for the last minute that we have. I do want to leave uh the viewers with a little bit of what you're working on, what you want to plug in, like where should people find you, like tell us a little bit about where you know, what you're working on and where can people get with you.

[00:18:24]So, uh I'm probably most active at LinkedIn and Twitter. I tweet about a a lot of my current thoughts and what I feel about the industry. I have also Swit uh club.com. Uh post I don't know the URL at the comments.

[00:18:40]>> it in the description. No worries. So, it's my blog with all my thoughts and also have there uh a lot of materials about what I think about agents and what I feel like from building agent in production. And also you should totally uh go and check out getkandlekeep.com.

[00:18:56]It's free. You have many books that I have been adding. Um I've been trying to create some kind of a publisher for books for agents. Uh and there are many free stuff you can try there and also get a free trial for 14 days. So, um go ahead and try that. And if you have any feedback, I would really love to hear that in any of the mediums you can find me. Um and that's it.

[00:19:23]Thank you so much Sal for coming. Uh for everyone watching, there's also a video about Kindle Keep uh independently, which I will probably post in the description. I posted it a few days ago.

[00:19:30]Um if you have any questions to Sal again, reach out to him directly over in his blog and commenting or in LinkedIn or wherever you can find him. Um and I'm just going to leave everyone with like one final thought. Um the people we bring on this channel are builders.

[00:19:42]They're building the next eight wave of the agentic age, if you will. If you are building anything interesting and you want to display it here on the channel, more be more than happy to have you on.

[00:19:52]Um use the email in my about section and hit me up. Thank you Sal for coming along.

[00:19:56]Thank you so much, Tom. It was a It was a pleasure. Likewise, everyone. Goodbye and have a great rest of your day.