ICM - Interpreted Context Methodology

Speaker 0: I want to talk about, Interpreted context methodology. This actually came from a guy that I found on Instagram, but I made it, you know, much more. This is much more. Oh, wow. It's a paper. Speaker 0: On on the archive.org, Sai. His name is Jake Van Cleef, but, holder structure as agent I have a strong conviction that, that you should never write code for your agents because, Interpretable, the next model will come out, the next harness will come out, and you'll have to rewrite it, and it will be out of date. So, what Jake Jake and I are very aligned on this, and I've now kind of bought into everything that Jake has said, which is, all of your agents, which is where we're going with with everything now, where the LLMs can do multiple steps, should be as folder structures. So, and I'll I'll get to showing it, but we can get to this little diagram down here. I can make this oh, no. Speaker 0: It's not even okay. I'll just do this. Make it bigger, bigger, bigger. But yeah. It's got this 5 these 5 layers where a claude.md at the top or you if you don't use Claude, it can be agents.md. Speaker 0: Kind of describes, everything that this agent it's it's the Sai. And and generally Sai feel this way about cloud dot m d files anyway is that it should be the signpost for everything that's within your code base or whatever you're trying to do. Don't try to stuff everything in there. But in this case it's your signpost at the top, and then your context dot, dot May is the where where do I go to find things within this thing as well. And then below that you have other folders and then it's a multi step folder process. Speaker 0: So you have your folders numbered. So you have in stage 1 you do this, in stage 2 you do this, in stage 3 you do To. And this will be a little bit, recognizable because, skill files are this way. So skill files are, I'm going to do a particular thing and here's all of the instructions I need to do it and here's all of the script files and helpers that I need to do it to. So it's just that Time multiple folders in a row, in a larger workflow. Speaker 0: So I feel strongly that, that everybody should be building agents this way. So the way that this looks is, the I originally started with the scripts and agents kind of top level folder that had all of my things and it's starting to, build toward this more this structure more, Speaker 1: see if I can make Speaker 0: this a little bigger here. Yeah. There we go. More a little structured now. So, this, excuse me, this content to guide folder here, used to be a skill file and I find myself more and more graduating from, skill.md files that do multi step processes. Speaker 0: I feel strongly that if your, skill dot MD file has more than 3 steps in it, it probably needs to be in this folder structure. So what that looks like is, this content to guide takes any any Context, let's say it's an archive.org paper or it's a To transcript or whatever. It detects what type Context you have and eventually it turns it into something. So it could turn it into an, in my case Sai put it on my Kindle in an easily readable format. It could also turn it into a LinkedIn post. Speaker 0: It could do whatever. But the point is that it has these, stages here. So detect and set up, extract the text, identify the chapters, split it, summarize the chapters, build a PDF, or build a Kindle. And there are, as you go along here, we can click in here and kind of look. Let me just shrink this down here. Speaker 0: Everything's marked down. Right? So, just like you shouldn't write code To define your agents you also shouldn't write code to define what your agents do. So you have these markdown files and it kind of just says whatever it's gonna do you know it's it's it's markdown, it's human readable. And the beauty of this is when you're when the when inevitably call it Opus 4.8 or 5 or what's the new 1 that's everyone's gonna Security's gonna Mythos. Speaker 0: Mythos comes out. Inevitably Time of these steps will be killed. Right? Because Mythos does it out of the out of the box. That's called the Uh-oh. Speaker 0: Inevitability. No. Good good. The Little lesson. Yes. Speaker 0: The bitter lesson. 2 points to Camp Geiger. It's called the bitter lesson which means that, there's this kind of old thing with machine learning generally where you could do specialize specialize specialize, the new technique comes out and they that kind of kills all your specialized things because the new one's just so much better. So, the great thing about this is you don't have to rewrite your code. You just kill the markdown file and Sai, well, nope. Speaker 0: Now the now the agent or the LLM just does that, automatically. So, running a little low on time here but you can start to see here where this folder structure, has both those stages there, but it also has Time skills. So the stages can reference the skills which can themselves be markdown files. You could also have Python scripts or other kind of scripts in here. You could have shared shared Context in here as well. Speaker 0: And then the other thing that's important and it's not shown in this 1 but is shown in a different 1 here in my meeting summarizer 1. Output. Output carries over. So you have an output folder and then this output will be the input for the Context, stage within your, ICM file. And I know I'm almost out of time here, but I have 2 more things, just to share. Speaker 0: I've put these out there and I can I don't know if we can do Time QR codes or something here? I know there's a way to do it within here, but but anyway, talk to me or Sai could send this out. I don't know if there's a way to send it out, but just some just a couple of skill files to either kick 1 of these things off where you describe what you want to do, what your flow is, to then build the ICM folder or this other 1 out here which is, skill To ICM converter. Because like I said I've built skills skills skills over time and I find that after a certain number of steps, Time saying 3 steps, it's it's a good time to then convert it into a folder structure. Any questions? Speaker 0: Answer your question? Speaker 1: Yeah. Speaker 0: You can put links on you can edit your, like, submission for the talk. Oh, okay. Cool. Then I will do that. Okay. Speaker 0: I'll do that. So I'll put these skills up there. Cool. I will do that, Rob. Thank you. Speaker 0: And if you don't get to this, then you could search Jake Jake Van Cleave, ICM Interpreted, interpreted context Methodology? Interpreted. Let me just make sure I got that. Yeah. Interpreted context methodology. Speaker 0: Yes. Go ahead. Did you have a question? Speaker 1: Where is this do you put, like, a visualization? Okay. I saw that there was a front matter there. Is there Time a specific output format that these MD files should be output to that other tools like consume? Time what is the MD structure that they prefer? Speaker 0: Yeah. At this point, Sai I don't think it really matters too much. Honestly, Time, you could read through the, the archive.org paper that he came out with, but I don't think the structure matters too much. The models are already at a point where they're smart enough to be able to, re read the front matter and read the steps themselves. It's all just markdown. Speaker 0: I think it's whatever works for you and actually as I walk through my ICM workspaces here I will notice some differences between them but the basic premise of I have some context, I have some stages, and I have some ways for those stages To, a guidepost to be able to point to Time different stages, is is the important part. Yes. Are there any benchmarks Speaker 1: for this approach yet? Speaker 0: Not that I've seen. I'm not sure. I will say 1 other thing, is that what you don't get with this and I wanted to bring this up too. I forgot. But, durable execution. Speaker 0: So there are, there are frameworks like temporal temporal, lang graph, all the all the ones that you know that you do write code for that give you the durable execution Sai you get deterministic replay, activity step boundaries, indepid I can never say that word, indepidacy. You know, logability, observability, all the stuff that comes along with those frameworks you don't get that with this. So I am currently trying to work to get my ICM framework which allows me to carry my agent context with me to also have these this kind of durability as well. Yeah. Speaker 1: So if we're switching the the order from, like, l usually, the code's calling l m, but does the Speaker 2: l m call code? Like like, To, like, do your workflows include a bash tool? Yeah. Speaker 1: That that that often translates from, like like, is that you Sai are you writing that as pseudo code that Speaker 2: Oh. Is meant to be transformed to to bash when LLC m runs? Like, is is that built into Time workflow or, like, is there a point where Speaker 0: It's more like it's more like progressive disclosure. Sai, like, the the, instruction file will have a pointer out to this is how I do a a particular thing that I need to do. So 1 of the things I do with, like, my meeting notes 1 is I, have it split large transcripts because Claude can only take transcripts who are so big. I don't know if it maybe it's better now, but, I have a Python file. So it's like when I need to do that, if I determine I need to do that, then I go and do that. Speaker 0: So Speaker 2: it's kinda So there's programmatic tool calls is not all happening Speaker 3: with direct, like May, single tool calls. Speaker 0: No. There it's all LLM deciding to to use those tools. Yes. Speaker 3: So I I I'm curious about, divergence when you go deeper, the structure. So you can have the higher level of abstraction, basically, grasping, certain that's that's what I should start thinking about. This is the world, and then this is what they're going to do. As you go down and refine, you can have some more variation, and that's okay. Maybe it's Yeah. Speaker 3: Doesn't impact the results. And I'm curious if you have seen something like that, like obtaining different results and how that would, basically relate to solid, all the way To, Time, 4.7 and Speaker 0: Yeah. That's a that's a good question. Yeah. That's a good answer questions. Sure. Speaker 0: Here you go. I'll just put that down, I guess. Yeah. I mean, LLMs are nondeterministic. Right? Speaker 0: So I think that's a definitely something to be aware of. I will admit that most of the ICM workflows that I have are personal and they're not mission critical either whereas, you know, if you are going to have something like that a more deterministic workflow like a lang graph or even just a Python script is probably gonna make a lot more sense and only leave things To the LLM that are truly, where an LLM needs to make that decision. Jake Van Cleef, the guy who came up with ICM, actually talks about this 60 30 10 split where 60% of the agent work that gets done is just like the plumbing and whatever Time connecting things together. The 30% Sai think I can't remember it was 2 Columbus then there was only like 10% where the LLC really needs to make a decision there. So I think as we as the framework gets refined or as you refine your own version of ICM, I think working toward that split would make more sense. Speaker 0: So Alright. Thank you. Thanks, David. Thank you.