Singularity Machines - Fast Software Services | Sovereign Compute Systems | Agentic Automation

Speaker 0: Good to go? Okay. We're gonna get started. Speaker 1: Alright. Thanks for having me. It's my first time here at AI Tinkerers. I'm from a company called Boston, and we're happy to be a sponsor. And my name is Edition, I'm a Tech engineer, and I'm gonna walk you through something called autonomous context delivery. Speaker 1: It doesn't exist. I made it up. But it's super important in, like, today's Agent. You just have to try shit. Lots to experiment Wirth. Speaker 1: And to build successful agents, will take you down a path that you won't really quite expect. So AI show of hands, how many people here have built an MCP server? Nice. How many people have made a Raj system? How about just an agent that ships to real people, real users out there in the wild? Speaker 1: Okay. And how many people here are founders? Story app. Alright. You're gonna have to combine all of these teaching, especially for the Founder, and I really want to call to attention that to build a successful business agent, going back to the basics is actually quite inter-agent, and you have to play the long game. Speaker 1: I'll explain what that means. These aren't real AI, by the way. I know AI Tinkerers discourages slides, but this is marked down in a code Edition, so it counts. Speaker 2: What is Posthog? We're a platform for Speaker 1: kind of everything. You've probably been familiar with a lot of these duct, things like product analytics, feature flags, experiments, error tracking, LLM observability AI traces, evaluations, data warehouse. Our main ethos is to try and ship things so that you can build a successful product solely off our platform. As great as that is, that made my job really, really hard. Why? Speaker 1: What did we build? My team is responsible for a particular agent called the wizard. It is a single line terminal command, n p x at post story slash Third, that spawns Ian ephemeral agent on your machine to do 1 job. Really empower. Install our products correctly. Speaker 1: But what does correctly mean? Oh, Code. It kicked our asses. So app the demo part, I'm going to run the command on the side. You'll see that it downloaded this nice terminal UI and spawns a Claude Agent SDK with some AI of like bells and whistles around detection of the framework that you're in the directory. Speaker 1: So if we authenticate, go back to the agents, what you get is an Agentic experience where no 1 has to go through the pain of integrating another manual SDK or think about what events to track. Obviously, it being an agent, it has access to your code BOS Ian therefore can infer things like what is your funnel, what is your checkout system, and instrument all the hard things for you AI what events to track, how to do error tracking, how to have session replay, all out of the box. And so you can see in the very top right hand corner with the tasks, it's auto populating tasks to be done. At the bottom, you can flip to the tailed logs of how the SDK is actually running Ian the Agents. Ian if you really Wafaa turn off your brain, you can just read New News. Speaker 1: The point of this though is that it's a very obvious kind of idea of trying to apply AI to certain parts of your business. Like how do I 10 x, AI, onboarding is this example. If I can 10 x my minimal base, that is monumental value. Turns out helping even a simple problem like this isn't so simple for 1 real Jason, is that any business problem that's worth solving never has 1 correct solution. So I'm gonna let the wizard run on the side. Speaker 1: At the end, we can look at the output. Who here recognizes this image right here, the long game? These little rocks. We got nerds in the audience. Love it. Speaker 1: This is a real time strategy computer Nate. Very nerdy. If you really suck at this game like me, the fun part is just fighting with Darji. But successful players do 1 very boring thing very Week. They expand their Ian. Speaker 1: Aune it's all about source material. And if you're building an agent that needs to solve a business use case, generative, the ceiling isn't at the harness or the context window. It's beyond the local optimization. It's still capped at the context level. It needs more source material, especially if it's a diverse business use case. Speaker 1: So what Jones a business use case look like? It might look something similar to ours or it might not, but I guarantee you the math gets scary very fast. So for Fast talk, we have 20 plus products, a ton of surface areas AI 20 plus SDKs, more frameworks, a ton of features, and this doesn't even account for what the user brings to the tape, their use cases, their requirements, their personas. So even if you don't have the platform we do, your combinatorial math is just gonna equate to something called a lot three Fast, and your agent has no chance to essentially navigate those situations gracefully unless you're prepared to give it a lot of context. So when I say the long game, there is very meaningful work to be done within the context window and how to optimize for things like progressive disclosure, but generally the ceiling that we've seen is at the context level behind it. Speaker 1: It's where you source knowledge. So before we Head the infrastructure to really solve this problem meaningfully, we would get very mean comments all the time on social media, like your wizard sucks, it's Bredun, my favorite comment is a hacker news or someone who Code opted our Agents to get free inference through us. To solve this, we Head to zoom all the way out. And so Fast thing I want to post use guys is think of context as an system, and that applies to your entire company, your entire business. Anywhere there is specific instructions or details that can make a human do a job better, your Agents would benefit from. Speaker 1: And that means being able to move context throughout all your systems. So this graph, which Guest AI looks like a weird pool table, is basically the very top right between the orange, purple, and green is Week have a monorepo that's over 5,000,000 lines of code for 20 plus products and tons of internal repositories Ian obviously documentation that tries to reflect that both internally and externally. We have to implement very large Raj Systems, and I know we have a speaker who's gonna have to talk about it, which I'm very excited about, where the knowledge graph has to constantly be maintained and created at all times. Ian, eventually, when that problem is can never be Director, but at least somewhat managed where your knowledge is being kept up to Date, has to be packaged into things like skills or resources that the agent can Compute. That's gonna require you to build APIs, releases, skills. Speaker 1: That doesn't even account for AI the time your agent is willing to do something worthwhile for a use. It then needs to execute via tools, MCPs, or LLC Nate if you want to subsidize the inference the way we So your context is part of a massive surface area that's as large as app your Compute, and if not Order layered. The first thing that we built to kind of, like, combine these disparate sources is something called the context mill. This particular diagram is basically a control service that we built team from the left to the right Jones assembly line of app soon as any engineer merges a PR for a feature update or any employee updates a company handbook, we can assemble that context, transform it, and package it with any set of custom instructions, example apps, anything, and create a manifest where we deliver that package to any agent on the fly in a couple of minutes. This requires a lot of piping. Speaker 1: But once you have something like this, this continuous context delivery lead you Sinay blow the lid off of what you can experiment with. So just to show what that looks AI, is in Head. Speaker 3: Where are we? Speaker 1: This is the context lead in action. We can generate thousands of library a library of thousands of scales just with simple build scripts. And once you have the build side of it all, you can start thinking about interesting introduces. Just like how App completely changed the way we thought about data retrieval and structured, we can do the same thing with context. So for what we did in the context Build is we have this kind of declarative spec where through YAML, we can basically write 15 different variants of an error tracking skill, for example, by framework Guest by pointing to the right combination of resources, just like recipes. Speaker 1: So all of a sudden, what would take, I don't know, a human weeks to generate 15 skills Ian be auto generated, presuming we've solved the source material knowledge layer of updating source code to documentation and then just making a composable interface on top of it. Once you have building blocks like an interface and a build pipeline, designed of like when you have APIs, you start seeing API endpoints in your data MA, you Sinay do the same thing with context. There are shapes that should exist bugs don't that would make for interesting combinations that no 1 would start from scratch. AI, for example, combining your privacy documentation with your pricing page is something no human would just think is a combination. It actually makes for a very powerful auditing skill. Speaker 1: And all of a sudden, we can have the wizard or any agent execute that by having these pipelines of building large libraries of skills in any combination that we want. Okay. So if things should exist, there are definitely things that should not. I think everyone here has heard of horror stories of what agents can do with 0 constraints. Because the wizard is Ian Assembly installed Claude Agent SDK, locking it down is paramount, and it jumping the fence is very, very scary. Speaker 1: It's at this point Week you have your agent now aware of any possible combination of context that it can be delivered is when you need to start thinking about the harness as a layer for consistent behavior and what you can enforce. And Head, we use basically anything at our disposal that's native hooks to the anthropic SDK. If you use Layer chain or the AI SDK, whatever the framework is, there are deterministic hooks and and ergonomics that you should use. But app the same time, there are infrastructure plays where you can use either the MCP Sandbox, but more importantly, things like gateway where if you are getting the inference AI and proxy through your platform, you can do much more critically important stuff to essentially ensure trustworthy and consistent behavior. I'm gonna talk and and show you guys something Week. Speaker 1: But if you have the open source repo, I encourage you to take a look. 1 example of having tool enforcement for something very, very, like, important is because the wizard runs on your machine, 1 of the big things to maintain a level of trust was to make sure that the AI is not reading your credentials Ian your environment variables. This is an instance where we Soalr we AI that using tools via an in memory MCP would allow a function to edit those files to have the AI be blind to its Tech. It's things like this where when you see the ergonomics of how the agent needs 2026, like, execute in the harness, it's a determine or have deterministic behavior, but it's context Ian the job to be done is something more fuzzy. So you gotta use uses both IDENTIFICATION if you're gonna execute something that's gonna be seen as both safe bugs valuable. Speaker 1: Once you have context being delivered in any combination Ian then you have the harness for, like, executing consistent behavior, the last step is feedback loops. What makes agents kind of like a new class of software is that it can improve itself. I'm sure many of you are aware of, like, this idea of compound engineering where you ask the agent to reflect on what it's done and what could have done better. We do the same thing. With every wizard Ian, and we have thousands of users running it every week, we ask it to write a little diary entry of what it could have done better, but from the context it was given Ian the 2026 that it was AI. Speaker 1: And all of a sudden, we have thousands of comments from itself in 1 long journal of how Week can improve itself. And then Fast PoC has a code editor that basically is married to all your product events and signals. I can ask it to open a PR on itself based off the observation it gave itself. And all of a sudden, we have an autonomous improving Agents from its own runs. Those 3 things Ian combination really only Services because we were actually solving a real business problem, which I encourage you all to do because business problems Karl, by definition, extremely diverse. Speaker 1: Ian if you combine those things in the right way, you'll move from those really mean comments to nice comments like these. It's something that has really shocked us even though we thought it was a good idea. We've basically 5 Tech our conversion rate from free tier to paid use, Lovable the amount Order double the time from converting to first event. Any objective AI metric that you would Ian a vacuum say that's an amazing business transformation was given to us by a lot of trial and error of building this agent through through the basics. So if we looked at now what the wizard's finished, it's going to ask you if you wanna install the MCP, give you a summary of the events that it instrumented. Speaker 1: This is a SaaS, repository with Stripe, Postgres, Workers instrumentation, all the things that you would Tech, and it basically inferred all these HINTS, and it ran via skills. So if we looked at If we look at the diffs, we have installed manage. We have the reverse proxy and the SDK initialized app well as event captures Salil customized to the user's code base. Combine these HINTS, and you can find amazing opportunities in your business. Thank you. Speaker 3: Yeah? We have time for questions. Yeah. Ian or 2 questions? Speaker 1: Sounds good. Speaker 3: Yes. Speaker 1: AI I would say we were pretty early, which, you know, everyone is because no 1 really has the best practice. It took us a year to get there, and we thought it would be super simple projects. Tech. Speaker 4: Mhmm. Speaker 1: Yeah. Ian. You're probably 1 of those people who left mean comments. We do not handle the monorepo well yet, and that's because we treated that as a deterministic problem of, like, IDENTIFICATION language and frameworks Aune MA is completely Code out of the water. That's where the next upgrade is to bring in reasoning via LLMs to make the decision there. Speaker 1: So look for an improvement. Yeah. Speaker 0: Mhmm. Yeah. Speaker 1: Aune amazing question. I wish we could talk about it for, like, hours, because it's Chowly about I think people are now starting to realize that your knowledge graph is something that is just, like, an inherent limiter on, like, what agents are capable of Coding. How do you have source of truth instructions? So for us, because we're a dev tool company, but this would apply to any Harness. Whatever your source material is, and for use, it's our source code that builds our products, We have both agent and human loops to have documentation be generated so that the knowledge graph is both Intensive, like, Jonathan expanding Aune maintained usually through a combination of Raj, just plain old PRs, and AI, Agentic processes. Speaker 1: But ultimately, our source code needs to produce accurate migration, and our business needs to then build meta Automation migration on top of that for any human, let alone Agent 2026 get, like, accurate use cases. Soalr, effectively, it's a very big AI, which I showed you in that graph. Cool. Speaker 3: Tech. So I just, set up a post hoc yesterday for my little app that I was Coding into this wizard. I did not use the wizard because it was an HTML only app, but I next time Team use it. Now that I know how cool it is. Speaker 5: Alright. Our next speaker, Speaker 1: I believe it's Raj, isn't it? Speaker 3: Raj, AI wanna come on up? You're already on the Zoom? Speaker 6: Week. I'm good with this. Speaker 3: AI I in? Main not a different Ian? Yeah. There was a little confusion about it. Speaker 6: Oh, I just copied a thing from HINTS. Speaker 3: AI. Uh-huh. 8000. 8833. What's the passcode? Speaker 3: 472. Yeah. 10. Yeah. Sorry. Speaker 3: I was on the other 1, I think. Oh, you need to click the join button. Sorry. Oh, you didn't click join. Yeah. Speaker 0: So I Speaker 3: should be Speaker 1: on there. Speaker 6: Okay. App. Speaker 3: Alright. Speaker 6: Alright. I'm Raj Bala, and I wanna show you something that I woke up building a few months ago. So 2 months ago, I woke up tracking, what exactly is Lovable? You know, we've all probably heard of it and seen it and kind of know what it is, but Lovable is a prompt to text creation thing. You type what you want, you hit go, and it'll build an app for you. Speaker 6: AI started thinking what actually makes up Lovable inter-agent, AI, what is it really doing behind the scenes? And I realized I had a lot of these components. I just needed to string them together and add some glue and I built this thing I Salil Perspect. And, I've been working on this thing actually for quite a while. It started off as this Sergeant app deployment teaching, and I realized that actually fit really well into what Lovable does. Speaker 6: So I Wafaa show you how I built it, how lovable is probably Build, and, tell you some of the hard things that I I ran into when I was doing all this Glue stuff. So I'm I'm gonna show you a demo of it first to sort of set the stage as to what the code, what the app, and so forth looks like. So if if if you're logged HINTS perspective, you'll see a chat window where where you can interact with agents and tell it what you want it to do. Soalr, the the idea is for people outside of this room, cheap who are not software developers 2026 be able to create apps and deploy them at high scale, create databases without having to know how to be a database administrator. So I heavy this existing app lead, and I'm gonna actually tell City, can you three the CSRF implementation and tell me how thorough it is. Speaker 6: Okay. So it's gonna do some things now. It's gonna talk to an agent on the fallback end Ian the Agents gonna say, okay, this AI looking for something that requires a software developers, so it's gonna farm it out to a software developer focused agent. That agent is gonna review it and and give me some feedback on what the CSRF implementation is. So I'm just gonna skip to that part. Speaker 6: So it's it's migrating the task, it's creating a a plan as all agents as all good agents do. It's it's data a plan Ian it's farmed it out 2026, sub agents. So but let me tell you actually what it what the pieces are behind the scenes. So there are a number of agents that make up Lovable probably and Perspect as well. So in my case, I figured, okay, well, I need a main agent that the user interacts with. Speaker 6: I need an orchestrator that takes all the work that needs to be delegated out Platforms it all out. I need an app developer agent Ian the app developer Agents got certain instructions and certain tools and certain capabilities, but then you might ask it, hey, can you create a blog post for me? So there's a content there's a content manager agent. Similarly, as I mentioned, the objective is to have non developers be able to deploy applications at high Soalr. But we all know that Agents, if you if you leave them unchecked AI Edwin was saying, you'll end up with a foot gun in a hurry. Speaker 6: Right? Agents unchecked will essentially shoot you. Right? And, so there is a, there's a security auditor agent that is required to review the code before it gets deployed. So this is to this is to prevent my mom from deploying an app that that might end up having a security hole. Speaker 6: So the security agent reviews all the Code, it's got best practices for what it might need to Ian. Ian and and as you might have guessed, all of these agents have a set of MCP tools that they call. So, some of the tools are things like app list file. So it Beads to be able to list all the files in the file system to see what you've got and see what needs to be created. So three are entire set of, MCP tools that are a lot like what Claude Code or Coding might use. Speaker 6: Right? If if you've ever paid attention to what codex is doing under the scenes, it's using a lot of UNIX type tools to list files and create directories and edit files. So so I use that that same sort of metaphor here too. In addition to being able to list files and create files and patch files and that sort of thing, a lovable like system has to be able to build and deploy an application. And this is where the the early beginnings of this thing that I built that, was essentially a React type of build and deploy system really came in in quite handy. Speaker 6: Solar, so this system inter-agent, when you go to build an app, it'll essentially build and deploy a React Router v 7 app. Team means anything to anyone here. So it's a standard app deployment model. The app developer here knows how to build and deploy React apps. It knows how to how to do that securely. Speaker 6: It knows where to do it. So that's the other thing that that lovable has to have is a place to build and deploy. So this was some of the glue I had to write. And and and it it it it it's perhaps 1 of the most interesting parts of this. So the agent came back and it said, oh, here's what your CSRF coverage looks like. Speaker 6: It's pretty thorough. It's broken. Okay. It's not unsurprising for my code at this stage, I suppose. But, so it gave me lots of good feedback. Speaker 6: But the the interesting part of this particular experience that I'm showing you right now is that all of this has to run-in a secure sandbox. So I can't just have the MCP tool Coding, and and executing untrusted Code. It has to run-in a sandbox. So what it does behind the scenes is it spins up a container Ian in that container is no private key. There's there's no secret that untrusted code might come into contact with. Speaker 6: So it's not like they can exfiltrate my secrets or they can do something to my service. So, so I had to build this thing. Luckily, a lot of Claude providers Hardware offering sandboxes today. So they're they're forking, container sandboxes. So I'm using a Sandbox from CloudFlare. Speaker 6: So it allows me to essentially create this MCP layer around the, the sandbox CEO sort of primitives and be able to do things like list file, create file, write file Ian so forth. The, other thing that I that I had to do that was pretty important was the security review. So I mentioned that, for you to deploy an app. So this particular app, it's saying, hey, the CSRF is is is not Guest, bugs there are gaps. So this application, if I told it to go deploy it right New. Speaker 6: So if I said, hey, can you deploy deploy this app? So it's gonna say, okay. Yeah. I'll I'll story to deploy it. It's gonna put it through security review Ian the security auditors gonna say, Nate. Speaker 6: This thing cannot go out the door. It's gonna stop it from from going out. And that's by design. But if it was able if if it passed, the security auditor would essentially stamp it with a unique Hirsch. And that hash is tied to that build and only that build could go out the door. Speaker 6: So essentially, if if you wanted to build a lovable like system, the title of this talk was how to build a lovable competitor. I think just to wrap it up, I think you need a few things. You need an Agents layer that that that you can spin up agents and sub agents and give them instructions and AI them to particular models and so forth. You need a a sandbox Layer. So a container sandbox like thing. Speaker 6: You're gonna need a build and deploy system that runs in that sandbox Ian you're gonna need 30to80 deploy target. So you could deploy it on any number of Layer. If you were to do if if you were to do three on my system, it would deploy them all on CloudFlare. But you could, of course, use AWS or DigitalOcean or 1 of those. So I'm being given the the time is up Main from the back. Speaker 6: So I'll stop three. Maybe take a couple of questions if possible. Is that okay? Yeah. Speaker 4: Any questions? Speaker 6: Thank you. Speaker 3: Oh, there's quite yeah. Speaker 4: Correct. Speaker 2: It's a good question. Yes. A good question. So the Speaker 6: question I'll repeat it. The question was this. It's like, why did you decide to to narrowly segregate what an agent can do? I've got, you know, 11 agents Ian they all have tools, but they all have different tools. They don't know how talk have the same tools. Speaker 6: So the question was, how did you Guest how did you decide to do this when it seems like the trend is to give, you know, to have all tools to all Agents? Is is that essentially what you're AI, gotcha. Gotcha. So so why did I decide to do a large number of tools rather than a small set of tools? Well, I took this approach saying that, hey, there are certain agent roles that need to do very particular jobs Ian I don't want them to do any other Jones. Speaker 6: But but there's a cost reason too. I I don't Wafaa load the context up with a bunch of tools and that just increases the token utilization team makes it slower. So for instance, the security agent has a very narrow set of tools and it cannot go outside of that scope. Ian the same thing with the content manage, the person that's writing the blog Fast, you don't need to give them the ability to build and deploy applications. They need like 3 Chowly, right? Speaker 6: Create blog Fast, publish, that's it, edit. Main sense? Was there another question here too? I thought Speaker 3: I saw another hand here. No? Okay. Yeah. Yeah. Speaker 4: So, I'll just do you have a test case for for a question. Yeah. Speaker 6: Use can. You can. Yeah. The question was AI can't you keep talking to the orchestrator when there are other agents doing Upwork. Yeah. Speaker 6: So it's it's it's an asynchronous teaching, Ian so, yeah, you can continue to to chat with the main agent or the orchestrator while other work is happening. Alright. Thank you so kindly. Thank Speaker 0: you. Speaker 3: Give give you this thing. AI. Is AI? Brian's here. K. Speaker 0: It's Week talk I've been waiting for. Oh, wow. St that's Tech expectations. So Guest share Speaker 7: from here? Speaker 3: Yep. You're already on the same. Speaker 0: Alright. Testing 1 2 3. Ian audio is gonna go through here. I'm not doing it from my computer at all. Speaker 2: Completion. Complimentary. I'll Speaker 0: talk about it. Speaker 3: So please welcome Ryan. Speaker 0: Thanks. I will start my timer here so I don't run over. Alright. So excited to talk to y'all. To beads or not to Beads? Speaker 0: That is the question, forking talk. How many people here manage projects and delegate Upwork in their projects to agents? Show of HINTS. Come on. Put them on. Speaker 0: Okay. How many people manage those projects and share information about those projects with markdown files? Speaker 2: Yeah. Speaker 0: Anything else that you manage those projects with? Beads. Beads. Uses. Great. Speaker 0: Anyone else Sinay Beads? What's that? QMD. QMD? Yeah. Speaker 0: What's QMD? Speaker 6: Like Apache Solar AI of like indexing of files Ian and Services on the moment for you. Yeah. Speaker 0: Anyone using MCP Workers 2026 Jira or to Limor? Local. Not a lot. Alright. So like mostly Wirth on AI, some MCP servers. Speaker 0: Okay. So, Ian AI advance tape like. If anyone wants to follow along, I'm gonna try and give some AI real practical tactical stuff and you can do it on your laptop as we Aune, and then I'll zoom out and give some of the high level context why I think this is super important. Actually, so Soalr I'll start with a little background. AI, have been so enthusiastically following Steve Yegi's Upwork. Speaker 0: Do people know about Gastown Ian then Gas City? Yeah, so he's super MA, use New, Week he started writing about Beads and the problems team he was having with MA files, it resonated so deeply with MA. Like I was using Claude for a good 12 common. I got pretty far in 2025 with Supercloud, projects, turning them into epics Ian then task, but you know inevitably the best case scenario with markdown files is you're just spending a lot of tokens and context on managing the markdown files. The worst case Sinay, which like clearly CJ was having Ian I had a lot of, I'm sure some people in this room have had, you New, the projects never Head. Speaker 0: Like the agents start revising the scope before they get to the end of the markdown files of the tasks. Does that resonate with anybody? Heavy that app? Yeah. Okay. Speaker 0: Soalr, you New, the trick to getting these projects right is always context and constraints. Right? You need to give enough context that the agents know what they're supposed to 2026, but then you need to give enough constraints that, you know, they don't meander and get off course. And Johanson who is using markdown files, I want to convince you to to take this home and try it tomorrow. If you start using Beads, you know, if your Ian is anything like mine has been, it is like magic the way you sort of can race right into the sweet spot of context and constraints by just changing the format with which you are talking to your agents about the projects you're trying to Compute. Speaker 0: Because with Beads, BDDS is an AI native issue tracking. And Soalr, you know, you've got a relational database, you heavy a relationship between like epics and tasks, and you can have epics that relate to epics that relate to task, use know, you can sort of, you have AI pretty big projects or pretty small projects. But then when you give a task to an agent, you know, it gets the context because it knows the parent and the AI, like it knows what has to be completion, workflows what the dependencies are. It can think intelligently about I AI Third, you New, it figures out intelligently about where the dependencies are Ian how 2026 parallel path Upwork, or what cannot be parallel data. And so you Guest sort of skip a lot of this stuff that you spend so much headache and tokens on with markdown files by going straight into Beads. Speaker 0: So I found that overnight AI prompts were just like so much better. And is anyone doing spec driven development? Okay. So, you know, I've been talking to cheap, so I've been doing these workshops about, three software factory intensive workshop. Went out to Seattle to join the people who started, you know, Boston and Gas City to do the first Main Tape. Speaker 0: Went to New York City this last month to do Ian, team month going out to California, it'd be awesome to do 1 here in Boston Ian you all want to do it. But tape there was a thread Ian I like, I lost the thread as I was telling you about the context and the constraints and I'm so sorry. Let's just Agent, I'm talking a lot. Let's get right into the example and I'll bring it back to software factories. Okay. Speaker 0: Soalr, whether or not you're using software factories, if you're working with markdown files, let me show you how Beads can make your life better. Okay. So, if you've installed Beads, let me pull up my terminal here. Okay. In this screen, I'm going to start a Claude session Ian Main this task, I'm just going to be in the, the repo that you may have cloned and there's basically nothing in here. Speaker 0: There's just a read MA file that says create beads for each of these. So lead just create some beads and work with them and see what it does. So I'm literally going to just copy and paste this whole thing into my terminal. So This says create a Beads for each of these Aune these are gonna be a few simple task, create a hello world script, make the hello world dynamic, right, replace world with a name, be able to use a list of names Ian make me a list of names I can use. I'm gonna come over Head. Speaker 0: I'm gonna jump into Claude, and I just pasted that. Should I make this bigger? Can y'all see well enough? A little bigger? Speaker 4: Yeah. A little bigger. K. Speaker 8: Good? Yep. K. Speaker 0: So while this is doing HINTS thing, I Build, you know, say another wonderful thing about Beads, it really is a master Claude Main agent experience design. So 1 of the things that if you read Steve's blog Fast, he talks about, like, following the paths that the agents would would follow author than trying to sort of bend them to his will, giving them tasks and then seeing what the agents do, and then when agents would try to do something over and over again, he'd try to figure out a UX design for the Agents, you New, to to do the command that they thought should Guest, or to follow the path, create the argument, create the functionality team they think should exist. So basically treating agent experience design the same way we treat user experience research and design activities. And so tape agents just get St. If you get tangled, you can ask Cloudflare Claude will just sort of figure it out. Speaker 0: Oh, you know what? AI, this is funny. I meant to AI really take it from the top AI from scratch with you AI, and so I blew away my Beads database. So look at this. I I meant to start by doing Beads Sinay, so what I should have done was BD init Ian if you're following along, do that in the repo that you cloned, that will create a beads database. Speaker 0: Apparently I don't even need to do that because when I pasted Ian, you New, create this with beads, it already knew what I was talking Boston. And then it went ahead and created tape beads for me. So here's a list of the 4 beads that it created and Ian gave these beads this AI really long Chief. That's okay. It's named after my project. Speaker 0: I would give it a Order, sweeter 1 if this was a real project. Go away a little pop up. Okay. If I like at this in the terminal, here's what this looks like. So I can see a list of all of my Beads. Speaker 0: It's like a list of issues in Jira, a list of issues in linear. If I do Beads ready, you can see that it only shows 1 item, not 4. So the reason is because it knows about dependencies. So I said make team a script and then update that script Ian then update the updated script, right, and then make a file for it. So it knows that there are dependencies. Speaker 0: If I do Bredun dependency tree, Ian I give it this Ian, Use can see that it knows the designed. So when I was talking about file it knows if it can work things in parallel or not, you know, it it knows the dependencies and it can organize the work accordingly. 2026 just demystify this a little bit more, like to show you guys how easy it is to use, I'll show, Bredun, and it's 1 of the links that I shared earlier. You don't need this, but if you prefer an interface sorry. That was the wrong command. Speaker 0: B d u I Soalr. Sorry. Teaching Fallback here. Soalr, open. Okay. Speaker 0: Here you go. So if stacks. Okay. Well, I'm not gonna make you watch me AI debug this live in front of you. If I ask lead to debug it, City file. Speaker 0: If you want like a Kanban inter-agent, like a, you know, a drag and Raj, you know, Kanban Karl interface, API d u I, which I showed the link to a moment ago, is the most popular community project, you know, if you want to get that kind of interface. Basically use can see, you can get this list here, you can get this list from Claude, you can say Beads, beads Fast, show me my Beads. And in the interest of time, I'll just show 1. Ian don't know. If you want to show, I can show you all of team. Speaker 0: But here, start working on Beads example, you New, 41 g. So now that I have my information Ian these, I can start saying to Claude, you know, work on this SPEAKER or work on the next Week or I can say Main up sub agents to work on as many of these Beads as you can concurrently. Or I can give it an epic with child epics, and overnight, I can say work this epic to completion. Don't stop until it's finished. I want the fully functional epic to review in the morning when I come back. Speaker 0: And the agents can find their way through these things. So, so I've said a lot of HINTS, I'm tempted to sort of zoom out and say how this fits into a bigger Layer project which is called Task City. Bugs let me pause and just 2026 if anyone has questions, comments, or anything before I do that. Boston. 1 more minute. Speaker 0: AI. Oh, awesome question. It installed the database called Duct. Dolt is like yeah. So the way they describe it is, if MySQL and three if MySQL and Git had a baby, that baby would be Dolt. Speaker 0: So it's version controlled MySQL. Dolt is a drop in replacement for MySQL, and it has all the Main, you know, AI you can Code Dolt branch and Duct, it's not Claude, it's Duct, something like you can create branches and and City has commits Ian you Main, like, it is a version control Date. And 1 of the things that's so awesome about it is if you Wafaa let your AI use on something and then actually know what it did and have auditability or be able to roll it back if something went sideways, you know, it can do all of those things. Yeah. Yes. Speaker 0: Oh, Aune you reminded me of the thing that I sort of lost the thread on with Layer. Thank you. Okay. So the Boston, if you have a spec, can you use that to generate the Beads too? Absolutely, so my favorite workflow here is AI generate an ADR, an architecture decision record. Speaker 0: So PRD, BRD, they're like lots of these things Main, but I always do an Ian, and so my first task will usually be helping me create an ADR, I'll create a Beads for that. Ian once I create the ADR, I'll point my agent at it and I'll say, now create a detailed implementation plan with an epic you New, that explains exactly how to implement what this ADR Ian, and when I'm satisfied with that, I'll say New create individual tasks for all of the steps in the implementation plan. And thank you for bringing us back to Spectrum and Development, a lot of people raise their hands on that 1. So I mentioned that I was doing these, software factory intensive workshops. A lot of people there are doing Spectrum and Developers, a lot of positive reviews for AI big projects with Spectrum and Developers, but a constant thing that I'm I'm hearing from people about that is it's awesome, but it really like Systems it feels like I shouldn't have to spend half of the day reviewing a spec, like when I just want to do a little a little thing. Speaker 0: Does that resonate with anybody? Yeah. Okay. So with Beads, you can give it just a 1 liner and say go, or a 1 liner and like task that three 1 liners and go, or it can be an epic that links to an ADR or a full blown use know work package from PoC kitty or whatever your preference is, Aune you can sort of AI size it that way. My hope about where this is all headed is that we will be able to take the best workflows of the PoC driven development frameworks Ian package them into different Beads formulas so that they Karl composable and mix and matchable with beads based projects. Speaker 0: Yeah. Yeah. By Third, do you mean the issues or the workflows Order Yeah. Alright. Check this out. Speaker 0: So I go Guest AI off. I'll show you after. Quick version is use can do Beads duct remote add and give it your Git remote Ian you can push the whole database up to GitHub or Duct or a different remote as well. Happy to talk more afterwards. I'm getting the task you. Speaker 3: Thank you. Okay. Our next speaker is Daniel Braden. Daniel, are you here? Yep. Speaker 3: He's gonna talk about Recursive harness builder. Speaker 4: Tech, Speaker 5: AI. Oh, nice. Cool. Okay. And what about if you just talk to it Ian then you send it another message AI a minute later? Speaker 5: Does it still Ian Sinay the whole conversation? Yeah. Because that's how the API is designed, which is silly. You have to resend the whole, like, hundreds of thousands of stuff every AI. But that's how it's designed. Speaker 5: So the way they actually leverage it is through caching. Right? The kv cache, it's kept warm for some time on the server. So if you if you're sending the same request, which the server is AI, I've seen this before. I'm not gonna charge you for the whole hundreds of thousands of tokens AI just seen. Speaker 5: So that's caching. So the idea of forking is that from the server's perspective, it doesn't know whether this new message arrived at the same chat that you just had Order it was a new Agentic is a local. And that's a fork. And I'm gonna just show, like, a little demo of what that would look like. So let's say that's our conversation. Speaker 5: I don't know if you can see, but there's, like, a user request, like, some tool calls, tool results, etcetera. It's AI a normal, like, prefix, of an Agent, just like context. Step number 1, we clone it into a different agent. At this point, it's AI IDENTIFICATION, the second agent can do HINTS own AI 2026 calls etcetera, whatever AI we keep the main Agents intact. Then the second agent AI that maybe I City a more complex task and does its own clone that has its own shared context. Speaker 5: At this point, we've paid once for this cheap. We've Main once for this prefix. So we can keep recursively doing this stuff or this agent can do its own neutral calls and report like a summary. And New, wow, you saved a lot of context Ian this guy can have another sub Agent and then they can report up Venue you can have another tree. Like, the point is there's really this is kind of like a context management idea because, AI, frameworks, for example, if you know Claude Code and Open Claude, for example, they have a thing that they nudge your agents to save memory AI, like, before the Tech Aune out. Speaker 5: Which AI, I think, it's absolutely stupid. There's absolutely no Presenter wouldn't they would pollute the main agent with this trivial task. I think they should make a clone of the agent-first the cache, do that sub agent, do the dirty work of whatever AI, building Murray files Ian report a summary. Head, AI made the memory files. Because I have not yet found a reason why they would not do that. Speaker 5: So if you know that reason, please tell me AI I'm going mad telling this to people Aune everyone's like, yeah. That makes sense. Why don't they do it? I don't know. So that's, like, the first thing that came to my mind that it's, like, an obvious optimization that should be everywhere in Snot. Speaker 5: There are a few things that are AI lead obvious that I think are worth kind of designing around. I didn't give this presentation some time ago, but I have like a brief list here. Yeah. Like there are like a few primitives of agents that I think are undervalued. File context length, obviously. Speaker 5: Right? We can't fit more Tech teaching in the context team it's designed for. Statelessness and limited reclaim, this is like a big 1. And that's kinda tied to the context. Like, the agent itself does not learn anything. Speaker 5: Everything it learns is in the context. So a human developer will figure out, like, oh, I had this bug 2 years ago. That's how I fixed it last time. Maybe I need to refactor. The Agents is not gonna have this. Speaker 5: The only way for an agent to have a thought like that is to actually have it search for similar bugs on every bugs. Have another agent potentially reason about that stuff. Like, it becomes that's that's the statelessness. That's limited recall. They can't recall things. Speaker 5: They can't learn three job. All the frontier labs can teach them all the math, but they're not gonna teach you them how, your project is set up Ian the architecture. You have to put it in files and you have to shout Main the Tech, which is limiting, but we just have to work around. I think it's an engineering challenge. And the other big 1 is, bounded intelligence. Speaker 5: And what I mean by this is, you know how you have AI Build files and you AI notice that the agent kind of doesn't follow them as much. Like, it Systems would mess up or forget things or get confused and stuff like that. And that's what I mean by bounded intelligence. Like, a single agent is usually good Nate, like, 1 Third, like, implementing 1 feature. It can do it reliably. Speaker 5: Once you start adding AI, like, conventions, tests, some, like, SSH, how to call your durable, the more stuff you layer, the worse it becomes because there's too many concerns and it's not really trained to do that very well. So from that was the idea for my project, which I Folder name, AI harness builder. Also, there's a GitHub. I'll I'll put it later anyone's. But, the point is to try to make a harness that is itself functions on the lowest level possible allows you 2026, first of all, make sub agents without a limited depth because Code Ian others usually allow you 30to80, like, have 1 layer of sub agents, but you can have sub agents who have sub agents. Speaker 5: AI. And the other thing is to allow forking because none of them allow forking out of the box. None of them allow your main agent to make a fork of itself as a sub agent. Soalr that's kind of I Build it in Upwork agent SDK. I'll have to refactor it because agent SDK, Claude agent SDK, it's not very good. Speaker 5: It's pretty bloated. And AI, I'm gonna actually do the demo of what it looks like. Actually, before that, I have a local, like, landscape of harnesses, to kinda explain and put it more Sinay perspective. This is just AI a quick idea of where different harnesses are in terms of we have, like, local 0 Ian that's, like, Guest API. Like, you Sinay this text, it returns Tech. Speaker 5: That's there's no harness. Then you have things AI Aune graphs, some, like, basic loops. Week heavy 2026 define tools. There's things like terminal agents like AI or some basic agent as the case where they already have some tools for you and it's Ian code so you can work from it. Then Fast of the stuff we use is Claude Code Aune codex and Hirsch and stuff with three feature Ricciardi well prepackaged harnesses. Speaker 5: And then there's also stuff like task down, which is not it's manage the agents themselves to do more complex work. So the, harness that I Build is more like low-level 2 that it's City generative. But what I want to show you is how I'm using that harness to build something Claude to level 4 with fewer markdown files. So let's show the actual thing. Real Week. Speaker 5: This is Telegram. If you guys don't know, Telegram is a messaging app. It's also the best the best, UI of any messaging app use Build find the best anything, Invest developer integration. So AI using it as a UI if that's strange. I could build a fancy website. Speaker 5: I think it doesn't make sense. I think Telegram UI is really really good. Here, for example, I have a lot of like projects which are group ChatGPT, in each of them AI have sub, AI, like individual agents which I talk to. And, what this actually looks like in practice is, I have a few So, this is just the harness AI. Now, the actual markdown procedures app they call them. Speaker 5: So, these are the files that the agents run. Basically, every agent starts with being injected this in the context and some like task that I give it ad hoc which could be like 20 minutes of me app them. Ian, this is pretty simple. This is pretty empower, AI, you're a router, you're going to work, whatever. Step number 1, step number 2, step number 3. Speaker 5: Ian, in this case, step number 1 is, spawn a scoping forking this tells you tells you exactly which Wirth parameters. So, it's literally like Soalr an agent with this prompt forking this procedure and that's just a different markdown file. This 1 has HINTS own, HINTS own CJ that it uses, AI Code the work, decompose the task, whatever, and it also inherits all the migration, because it's a forking, which, works remarkably well. So, that's kind of the idea, that Ian this case, this router creates a scope and the scope returns some like, if it's a empower like complex task Ian as if it's a complex task, then it tells it the router to spawn another router. And AI have a little graph here AI showed some time ago that this AI of what it looks like. Speaker 5: And this I'm just migrating what the procedures look like that the router procedure creates the scope procedure and the scope procedure decides like this is not a simple task according to the criteria that I designed. It's like many files, many stuff so it explains it down 30to80 three. In a new router then, makes a new scope and they just keep decomposing and looping and Coding down, down Ian down until it's simple enough and then there is like an executor file and then a verifier file Ian then they keep forking here. Executor AI, I'm Aune. Verifier is like, New, you're not Aune. Speaker 5: And then they come back and then, and then this results in like three, very long agents three. AI, that's an example of 200 ish agents. I have trees of like over 1000 right now Ian to be clear, these are agent-first unquote. Most of them are forks of each other. So Fast of them no. Speaker 5: All of them are forks of each other actually Ian this Nate, because I think this was just 1 Nate. Yeah. So LLC of them are forks of each other. So and Fast of the league, AI guess, 1 or 2 or 3 agents Main be running in parallel. So it's 400 agents, but it's kind of like Ian very well context managed agent. Speaker 5: And AI have a few more, for example, here, I have like, I can run a tree command which is gonna show me, like, the current tree of this project, which is, yeah, which is, the actual agents working on themselves. And, here, you can see how, like, each line here is a different agent. So AI would, like, click here and I see what the hell this guy was doing. Use. I'm I'm just gonna wrap up Wirth, like, some conclusions of, the point with this is that, I found this very powerful because I end up having the agents work AI non story for 24 hours on Claude their own task, decide on priorities, AI, anything that is doable by an Agents, I can force them to do City. Speaker 5: Because cheap it's AI an ambiguous goal, they end up debating what's the original what's the actual goal is. If there is, like, no verification loop or there's a AI verification loop then they can verify it. And talk of this is done with these AI 5 markdown files. And AI think it's pretty powerful because this has no memory system. There's 5 there's no line graph. Speaker 5: There's no there's no code in this emergent behavior Ian yet it produces like a tree that delivers a very AI task down quality of Presenter. If you know task down, except without the bloat of the gas down duct with like AI procedures. And this AI back to my idea that I think Event beads being very agent friendly AI be a little bit of a, like, we're still trying to force the agents into their narrow structured outputs. It has to be this issue that's blocking that versus in text Week they communicate to each other, they just AI private a AI a AI a reports Ian then short instructions Ian then they can introduce not like this issue is blocking that, but they're gonna be AI, this issue is blocking that, but there's a priority, but there's a constraint Ian this unstructured text, they're really good at it. And I think there is no reason to constrain them to structure if they if we don't have to. Speaker 5: And that's why I still am a believer of markdown files. Aune, anyways, if any of that is of interest to you, connect to me on LinkedIn. I'm not sure how long I'm gonna be staying for today, but shoot me a text or find me here later on. And you you can also try out my project as well. Thank you. Speaker 2: Daniel. Speaker 3: Yeah. SPEAKER. Is Jason. Jason. Yeah. Speaker 5: No worries. You've heard it. You've heard it before on my Speaker 0: yeah. Yeah. Yeah. Speaker 4: Jason is Speaker 3: gonna talk about Singularity Machines, fast software services, software and compute systems, agentic automation. I don't know what that is, but it sounds interesting. Speaker 2: Is it sharing screen? Speaker 1: Yeah. But I can't play the sound there? Speaker 3: Daniel or Daniel? Yeah. Speaker 0: Yeah. Yeah. I'm not getting them, Speaker 4: but I'm team. Speaker 0: Use go again. Right? Speaker 2: Yeah. It's not getting I Speaker 0: don't know what's app. Let's see CEO. Speaker 2: Okay. So share. K. It should be sharing now. Okay. Speaker 2: There we go now. Alright. So I'm going 2026, it should be sound. Let's Speaker 0: see. Speaker's better. Task you. Okay. Speaker 2: Okay. Can we try the HDMI cord? Speaker 3: Oh, you need to, oh, I see here. May tape to disconnect from the Zoom audio. Otherwise, we're gonna hear. Speaker 2: No. Speaker 8: I don't see Speaker 0: Are you good? Speaker 2: It's it's a little slow and nonreactive, but that's okay. Speaker 0: Really? Okay. Looks like AI Speaker 2: go like this Main so everyone can kinda see it Claude there's a voice agent 2026, so it's gonna be kinda tough. Speaker 0: Yeah. Okay. Speaker 4: Week, okay. Speaker 2: Let's just start it up and do the best we can with what we have to work with here. Speaker 0: Okay. Speaker 2: Okay. Time is starting right now. I team my Jones, which is right. Okay. So I'm just gonna start off with the origin story really quick before we get into my demo. Speaker 2: So origin story, my first AI Tinkerers event was in November 2024 when I demoed Manage Master Code challenge coach and interview prep agent. So now I'm pretty much calling everything an agent. Then I met regulars from Sunday Claude, later joined their group as a regular, and since then have attended 39 AI stacks over there over 2 years. Now I find myself here again 2 years later with an even stronger addiction for building and presenting migration. This has been an amazing journey so Karl, and thank you to everyone involved Ian organizing Main both three groups as their ground 0 for practical AI building in Boston. Speaker 2: So thank you guys for Ian, and thank you to Sinay Club. If you don't know about Sunday Club, definitely check them out. Revealed great group. Okay. So the company that I'm launching is called Singularity Machines. Speaker 2: So Singularity Machines is fast software services or speedy code changes and testing in any Tech stack. We also have we're we're building sovereign compute systems or sovereign systems from the hardware up using machine code, assembly language, and mathematical programming language, and a local private AI on the top to control the whole thing. And then the third part of this company is gonna be agentic automation or harnessing your off offline models for desired agency. So the building Ian preparation for Singularity Machines comes in the form of building and launching AI agents that cover each aspect of building the company. So if we minimize Speaker 3: this, I could take give you a Speaker 2: look really quick. So this is the experience section of my LinkedIn profile. This is where I'm posting all the agents that work for my company. So if you look Head, we have the attention agent, which is for marketing, Ian book 9, which is a human translation Agentic AI puzzle agent, which is to help you for your coding interviews, quantum program Mathematical Agents, how to, you know, how to learn quantum programming from a mathematical perspective. And then we have a local private offline voice agent, which you see on the right side here. Speaker 2: So if you go on my LinkedIn, you could see all the different agents that I'm working on over time. Let's minimize this. And what I'm gonna do here is I'm gonna bring up so now we're gonna take a look at our voice agent. Okay? We have the tech stack on the lead, Speaker 3: so you Code take a look at Speaker 2: all the tools that we use. There's actually 2 more pages that show the rest of the tools that we use, and let's just see how this local model works. So it's using Bala 3.21 b, which is a local offline model. Okay? So we're not using any API. Speaker 2: No credits. No tokens. Totally free. Ian, actually, this is the the main point of this company is to kind of give people the power to, operate totally offline-first if the AI is powerful enough, it should be able to do everything for you in an offline environment. So we're also attacking privacy with this company. Speaker 2: Okay. So let's try the voice agent here. Hopefully, you can hear it. Let's see what happens. What is your name? Speaker 7: I'm Singularity Machines Nate. Speaker 2: Okay. Why did it say Terminator? Because I'm trying to show you how delicate the harnesses are. The only harness I gave to this agent was the name Terminator. Okay? Speaker 2: So now it's assuming all the roles of that from the movie. What is your main purpose? Speaker 7: I'm here to assist you with information Aune access. What can I help you with today? Speaker 2: There is a big group of people watching you. What can you tell them about your service to humanity? Speaker 7: Good afternoon. I am here to assist and provide information to the public in a neutral and respectful manner. My My purpose is to offer accurate Nate countable responses, answer questions to the best of my abilities, and engage in productive conversations. Please feel free to ask me anything, and I would respond accordingly. Speaker 2: Are you on human side or the robot's side? Speaker 7: I'm not with either party. Information, not to take sides Ian conflicts. I strive to remain neutral and impartial. Speaker 2: Okay. So if you take a look here, we have a bunch of team tools. It gives you everything from the local AI. We uses llama llama, local speech to text, whisper, whisper CLI. There's a bunch of different Chowly. Speaker 2: You know, every time you build something new, it's gonna come up with a brand new tech stack. So I think this day and age, you need to be able to handle all these tools that are thrown at human. That's the beauty of AI. That's what it's unlocking for all of us. So we have more tech stack information over Harness, and then I just wanna highlight this last slide here. Speaker 2: So when I do my AI Code, what I do is I'll tape so as I'm coding along, I'm verifying and checking and testing so I don't have to backtrack later on. And Main that way, you can build a Tech level app working up to the top. So here you heavy, actually given the common. So this block is the next bridge between the app and the offline local model. Reviewing this block line by line before merging is good engineering practice because model calls, endpoints, request shape, and message roles are high impact migration points. Speaker 2: Careful review here prevents avoidable bugs, reduces backtracking, and makes the code BOS easier to maintain. Okay? So here we have we have the main, function right here and just gives a little description of what's going on. If you really Wafaa learn the syntax of the code of the Tech stack that you're using, take 10 minutes and type out the code yourself. Start to burn it into your memory. Speaker 2: Okay. So I Wafaa show you 1 more agent that I'm working on which is called Ian book 9 agent. Here is the Tech stack for you. The first part of the text stack which shows the front end. Okay. Speaker 2: So 1 book across many Language, 1 book 9. So my I was an English teacher all over the world for many Beads, AI 2026 years. I taught Main international schools and I also studied languages on my own. I actually story 9 languages on a weekly basis. Okay? Speaker 2: So this is my attempt at helping me continue to learn the languages that I've learned over time. So what you see here, 1 of the best ways to learn a language is to read the same book in multiple languages. So if you're learning Spanish and Italian, read the same book in both languages Ian, going through all the differences will help you strengthen your understanding in both. Okay. So here we have Wizard of Oz. Speaker 2: Right? So the beauty of this is it does the character voices for each guarantee. It does all the translations for use, and it's all set up. So when you're learning a language and you Wafaa listen to the book, you need to distinguish between you need to distinguish between the characters in the book. So I gave each character a voice using, we'll get to the Tech stack after. Speaker 2: So let's take a look at the English version. Listen to what it says Ian then also listen to the character differences. Uses Speaker 3: Dorothy. Toto was a fine curiosity to all the people, for for they had never seen a dog before. Speaker 7: How far is it to the Emerald City? Speaker 9: I do not know, for I have never been there. It is better for people to keep away from Oz unless they have business with him. Speaker 2: Better to keep away from Oz unless he has business. Okay. Let's listen to the Spanish version at the same time. You should be able to make out the different character changes. The translations are working fine. Speaker 2: They work great across the 3. Alright. Soalr this is this is our Automation, human translation agent Ian actually this is feeding into the Sovereign compute systems because I'm approaching it from a translation perspective. So use have consciousness, translates down to human Language, which translates down to all these machine languages and all the way down to the hardware. So I'm approaching this new design from a linguistic perspective. Speaker 2: We have some more Tech stacks stuff over here. Let's see where we got the local language processing. So Ollama runs local a on models on your machine. Deconstructs handles local interpretation, structuring, translation, and JSON generative. Common 3 3 b serves as a lightweight helper model in the project config. Speaker 2: Llama 3.21 b serves as a smaller utility model in the project config. Python Guest sends prompts from Python scripts to the local Solar API, etcetera, etcetera. So when I do this stuff, I like to really take a look at the Tech stack and see, you know, what I'm using, etcetera. And, I think that is pretty much it for my presentation. Anyone has any questions? Speaker 2: Thank you. Yes, sir. Speaker 4: App. Okay. Right. Speaker 2: I basically so my approach is to start with the smallest, crappiest model to begin with. Just run through City, use it, use it, use it until it starts hallucinating, until it starts having those big issue, and then I'll just switch the model pretty much. So if it's not up to the task, just switch. But right now forking something like this, it's it's pretty much okay as a voice agent. You could beef up the harnessing, and I'm sure a very small LLC model like that. Speaker 2: If that model is crappy enough, you Code probably find Order powerful small MA. Like, I think my next iteration of all this author I factory reset my whole computer, I'll be using the like AI offline models because I wanna go through a series of those and see how they perform. But in general, I'd say forget all the important benchmarks out there and just use your own usage as the Guest benchmark around. Yeah. Soalr try them out, dump them if they're bad, Guest a new 1. Speaker 2: Yeah. Speaker 3: What are you using for your your? Speaker 2: I think it's Guest, it's just a right Head. It's just a little yeah. It's just a little sentence in the the Code AI, which is right here. Ian me see. SMT 4, this 1. Speaker 2: AI, like, yeah, it says it right here. You are Singularity machines terminator and AI local sentence. Speaker 4: Yeah. Speaker 2: Thank you very much. Thank you Speaker 4: very much. Thank you very much. The Speaker 3: next speaker Speaker 1: Thanks, Trevor. Speaker 3: Terry. Hi. That's not Hey, Eric. There you go. Ian, Ian's talk is called the duct tape was hiding 3 bugs an Event Speaker 8: Migration Story. Speaker 4: Outside of, outside of, of the various cards that I had in place in order to ensure that it's doing new, new for Anthropic or a new tool or just new in the AI and, agent space. It's it's a lot to keep on top of. So we need to put in a lot of time, put in a little stuff back out of the guardrails. And then while trying to New, like, there's gotta be a better way of doing this than going and searching and finding out that, like, that's the right thing to drop it, actually, created this thing. That's just called structured output. Speaker 4: They guarantee a server side that your JSON's gonna have to be backstartsheet data. Okay. Perfect. So we're not everything that we did, put that in, and as soon as I did, all of these bugs that were hiding came to light. So the very interesting thing about it is that we'll be able to pick up, actually, if you don't know time at all, go back to get into that. Speaker 4: And because that's also now part of the very robust unit test, maybe I can tell a little bit more to hear that, but, I think data model. And but but they're there now, and they're able to so they're both here to instantly quite Speaker 0: Peter. Speaker 3: For our last and final SPEAKER, Earl. Hello? Speaker 4: I can pass it out the nest, Speaker 3: and you're kinda spilling the air around. Speaker 8: Can you hello. I'm sharing. Can you see it? Speaker 4: Okay. Cool. Speaker 8: Alright. Great. Last but, hopefully Nate lead, good to be talking with you guys. I'm Karl, and I'm, working on app project called Nimbalyst. Speaker 3: Lead me just, show you the Speaker 8: GitHub for that real quick here. And it's a visual workspace for building with codex Ian Claude code. So it's, we live in it. There's about tons of people, like 10,000 or Soalr Date users now hiding in Nimbalyst. And you could think of it app, like, a cursor competitor, Code codex, Claude Coding, but we have grand vision. Speaker 8: So it's actually a unification of a local linear, Ian. So MA sharing, Excalidra, Salil the like together with, Agentic AI, plus a bunch more. So lots going on there. I'm not gonna demo all of that. I wanna particularly focus on this harness concept, which a lot of people are talking about. Speaker 8: 1 thing we realized over time app we've been building New is Ian a sense, it is our harness for developing. It we have a kind of expansive view of harness. City the the the actual workspace is part of of the harness as Week, and I'll I'll make that case a little bit. So let me show AI think I'm gonna go in, light mode because I think it's a little easier to see in the back of the screen. Okay. Speaker 8: So this is what I wanna talk with you about is, this notion of the harness. And so Week spent about, I don't know, 30% of our time investing in this. And the more we Invest in it, it compounds because now we can actually have our agents work much better. So the more work we do on the harness, the better our team gets and the faster we go. So you've got your MA, you've got your Agents, which are harnesses themselves app others have talked about. Speaker 8: And this is what we're AI of thinking about is this this harness around our development process. And I have 8 pillars. Many of them have already been talked about today, so I'll go through a couple, deeply and most of them quickly. So that's what I wanted to show you. So, let's jump into it. Speaker 8: So all of our learning on this came from failure modes, basically. Week fail at something. Tape agent would fail at something. We revealed, oh, we gotta improve that. Let's let's go make like it better. Speaker 8: And, and so get some of these screens out of my way. AI. So oh, and let me the example I wanna do here is suppose I'm working my product Ian I decide I, I wanna make a slightly better, feature here. So this is, our Boston Ian I got my little, pills Head. And I wanna when you hover, I wanna say, you know, how many other items have the same Build? Speaker 8: That's the new feature I wanna build. So, like, hold that in mind. How are we gonna have a harness that makes that happen better? So the first pillar of our harness is context. Everyone's talking about that, of course. Speaker 8: And here's AI in my Claude MA. And, 1 of the things that we have on it that might be useful for you guys to implement if you don't is the documentation reference. So PoC picks up the Claude, the agent uses, Ian it's like, okay. I'm gonna be building this little UI widget. What what is relevant to me? Speaker 8: So we actually have a document on floating UIs because Week it's a UI duct. And so it might read that and might realize, okay, never manually calculate the position. Or we have, you know, I three IPC listeners are gonna be relevant Event, the Jotai AI. Use know, so it's gonna read those different documents. We never call, you know, all sorts of problems with IP listeners. Speaker 8: So we've wrote them all down in here and it's gotten a lot better at that. Right? So writing down our learnings, putting that Nate Claude MA Ian it picks that up for its context. 1 thing revealed to that is we we have a diary of mistakes. Others have been talking about mistakes. Speaker 8: When you get a mistake, it goes into the diary. If it happens enough, then a Agents written on it Ian then it goes into the CloudMD. That's sort of our process on that context. So that's the first file is Tech, much more to to say there. The second Week call private. Speaker 8: You could also call it, like, a context graph, and there's sort of 2 layers to that. So, 1 way to think about it is when you're making even this very little feature, it's Nate, it's not just a prompt to do this feature well. Use already showed you a mock up that I have on it. We write a little PoC. You might do Director of Ian development. Speaker 8: You might not you know, we write a little plan. Maybe it's 5 paragraph Guest 2026 sort of set it up better and memorialize what that is. We might have, a bunch of sessions related to it too. And even for something smaller like this, it grows. And how do you keep track of all of that? Speaker 8: Right? So we think of that as the context graph of the private, and there's sort of 2 ways to think about it. So 1 would be in a local, combine kind of teaching. You actually just write it down. Or if you're using linear or whatever, you just write it there. Speaker 8: So like making the links between the plan document, the mock up, use know, the the the tape, playwright spec you're gonna do test driven development on this. That's everybody could do. A feature of Nimbalyst we sort of invested in because we AI how valuable it was was, okay. Let's actually also link all of the sessions, together with it as well. So I can have the agent come and read all of that, map it together. Speaker 8: And so it's so easy for me to come back and say, remember this thing we put on Folder Order pick up everything related to City. It's all in the context graph and I go. So I could jump into 1 of these here and, jump into this this particular session. Now I've got all the sessions that are related to it together, and I've got the files that are related to it as well, and I can jump into those files. So that's the second big idea is to kind of context grab and build it as much as you can. Speaker 8: The third big idea is a pillar AI of of our harnesses is capability. I'm not gonna talk a lot about three. Everyone's talk about this Build Chowly specific to your use case. Of course, you got tons Ian Claude. I'll just show you a Week, quick example of that here. Speaker 8: Ian AI the right 1. I think it's here. Yeah. So I've got, yeah, I've got these different tools that I gave it to read AI process logs, to look at my database, to check the renderer. We invest a lot in those tools so that it can actually be using the software itself and, and gaining the capability to understand our own software and build on it. Speaker 8: Fourth pillar is workflow. So building, and there again, I'm Speaker 3: not gonna spend a lot of time. Speaker 8: We're all very familiar with these teaching. Building the skills, building the the, the the slash commands, whatever. We do a lot of investment to make sure that they're cross agent because I don't wanna be locked HINTS Claude code Order codex or the AI. So all of ours are cross Aune Jason agent and kind of structured away so that we can just use those workflows however we want. Alright. Speaker 8: Fifth idea is restraint. Again, we'll spend a ton of time on that 1. Order people have already talked about that Date. But, building Ian, heavy. You cannot commit to Main, you know, permission for that. Speaker 8: You can't add another database to our product, which is how I was trying to do, things like this. You can't restart the app without without asking me. And Week do a lot visually app you kinda tell use. 1 of our thesis is three contrary to the terminal kind of, approach. Of course, you can use a terminal Ian Head, but, you know, humans are visual. Speaker 8: Show me the information as visually as possible. And so let's just look at an example Head. AI I have yeah. And so, you know, kind of a visual common. Before you go and do that commit, what are you gonna do? Speaker 8: Modify it, commit. And we have lots of visual check ins where the person can actually manage and monitor this. I'll return to this theme in my last completion points, but I think that the harness I keep expanding the meaning of harness, but the harness actually includes the human and agent interaction. That's part of what's gonna make this agent effective. So for us, that's a lot visual. Speaker 8: Alright. So fifth was restraint. The sixth 1 is verification. And so, I think everyone's doing that. You're not you should be. Speaker 8: So, being able to tell the agent to go and run, play AI Ian our Nate, you know, for front end stuff, do test driven development. We always dreamed of it over the past 20 years. Now we're finally actually doing it totally and completely, especially for the night runs where we just go and build a whole bunch of stuff and it loops and loops and loops and verifies, against the, the spec. Alright. So the last 2 I wanna spend a little bit more so that's a verification. Speaker 8: That was the sixth 1. The seventh is, and the eighth are the 2 more human related 1. So that that Build the harness for the human and agent to interact with each other. And so, 1 of those is just Build a visual inter-agent. And and let me just demo that for a second here. Speaker 8: So I come in Aune I'm I'm in this, work stream or work tree AI to build this tag, 2026 tape, and they just made a whole bunch of changes. Well, I wanna be involved in that. I don't wanna jump over to Obsidian. I don't wanna also pull pull up something else. So I just pull pull it up right here, and I can see the, this is my spec for it. Speaker 8: I had it go and make some suggestions and changes. If if, if those let's like at them Head, you know, AI, a change in in the in the, mock app, AI can everything is red, green, different. I can kind of look at that and approve it or not and step through the changes. And so I'm looking at my files together with my sessions. That visual interface is key to our harness working effectively because I can manage my agent better. Speaker 8: And then the last point related to that is, Fast pillar really is coordination. We need to be able to coordinate the work across multiple agents. We really use our job of this harness is to enable, you know, 10 agents to be working guarantee, simultaneously. You see a little bit of that over here. But then, of course, we have app common board, and you can jump in and review and manage and filter and Aune, and the AI, both for your, sessions, but also for your, all of your tasks and to BOS and the like as well. Speaker 8: So those are the builders that I Head, and I think that's what I Wafaa show you. And so jumping back to this, by putting these together, we found that, you know, some of the principles are, own your own harness, invest in your harness, iteratively improve it, make it a harness for the human and for the agents together. And in so doing, we've been able to build, you know, just a massive amount of software at a much, much higher quality than just letting the agents run on their own. That's it. Speaker 3: Questions? Speaker 4: Yeah. I have a question. I have a question about your product, which Speaker 8: Yeah. Speaker 4: From the least picture that Speaker 5: you showed, it looked Speaker 4: a bit like a jump table for basically progressive context discovery and, like, when this happens with this file, when that happens with data. Why is it not a set of skill? Why did you just make it on a keypad? Speaker 8: Well, because we want that each of those that it's jumping to is is a document on, a piece of the architecture and how to think about that piece of the architecture. So in the end, it's a MA AI, and we found that that worked a lot better than having it be interpreted as a Build, 1. And 2, we're we're very cross Agent. And so the way in which codex is doing its skills versus Claude code versus open code versus whoever, by just having a straight markdown, it was able to work better cross agent. Yeah. Speaker 8: Yeah. Speaker 4: So for for the slot integration, are you guys, like, dependent on clock dash Yes. I know. Yep. Speaker 8: We're all we we're just taking all of our tokens for the past 2 months to make our codex integration and open code integration amazing. And it's now better. I mean, it's now really good Ian, you know, so but yeah. You New? It's unfortunate. Speaker 8: Yeah. Speaker 4: So this seems to put a lot Ian the context. Speaker 8: Yes. Speaker 4: For, you know, presumably for BetterHelp, but what Jones it how uses that impact your, I don't know, your your flow of walking up to start with the Edition, how much you can get Aune between session in 1 Edition? Speaker 8: Well, Week we Wirth very multi-session. So we're we're running 10, 20 sessions at the same time. Right? Different work streams Salil all in this tool. And, so we we haven't found City Director, like, our mental workflows, from from the agent itself. Speaker 8: You know, you just Ian City as separate as we're building up each separate session. You know, the only concern that we've been worried about is what about burning Tech? You know, how much are we putting Ian team of the token burner? Week managed that way down. I think Speaker 4: AI sort of written on the Speaker 8: end. Yeah. Exactly. Yeah. Opposite of Kingman talk. Speaker 8: It is. Yeah. Yeah. I mean, we're spending more tokens to get better results because of the Layer AI that. Yeah. Speaker 8: You Code use it for anything. We're focused on product manage Software just to stay focused. But we do have like, we have a lot of users Aune we have writers using it. We have, like, CEO types using it for CEO stuff. You know, we have, there there is a diversity of a lot of marketing people doing marketing stuff, you New, like because it is a general workshop, but and it's actually there's a whole extension system, so you can build an extension for whatever you want that's natively integrated Aune agent native. Speaker 8: But and so people have built some really cool things, you know, but I Week have to stay focused. So it's open-source. Go do whatever you want, but that's what we're trying 2026. Alright. Thanks. Speaker 3: Thanks, Baragal. Alright. So we Week at the 80 third and Zoom app. We could inter-agent around a a lightning round for people who can just drop their phone and just talk for, like, 3 seconds, or we could use the next step. Can I get a show of hands? Speaker 3: How many people would like to do lightning round? How many people would like to network? Yeah. I think we've got more for the networking. But if you wanna walk around and do your own version of Speaker 2: okay.