Jitpal Kocher on why Context Rules AI - Episode 2

[00:00:00]

Introduction

[00:00:00] Jakob Heuser: Welcome to The Task at Hand. A podcast focused on the builders and makers putting great things into the world.

[00:00:06] ​

[00:00:12] Jakob Heuser: Every team using AI right now is quietly becoming a context manager. paste the same brand guidelines into every chat. You hope your teammate went to the wiki, downloaded the latest PRD, and pasted that into their Claude md. You rebuild the same mental scaffolding, whether it's Claude Cursor, whatever came out last week, you're pasting it everywhere. The models keep getting smarter, and yet somehow your workflow keeps getting dumber. Today's guest thinks that the problem isn't the models, it's that the context never had a real home. JP is the founder of wire composable context containers for AI agents, he's been arguing that context engineering is the primitive, most teams are underestimating. He's also a 10 year product growth veteran, a fractional operator, food writer and fan, and loves the NBA in roughly that order, depending on the day. welcome in.

[00:01:09] JP: Happy to be here. Thanks for having me.

JP's Background: Builder from the Start

[00:01:12] Jakob Heuser: Yeah, not a problem. It's, it's always fun to talk to people when they get, when they're talking about their products and the things they love, and you've got a lot of different stuff that you love. So, we could probably just spend an hour talking about NBA stuff. We're here to talk about wire, and kind of how you got there.

[00:01:28] So obviously Wire was a more recent product

[00:01:31] JP: It is.

[00:01:32] Jakob Heuser: and most of your career growth was actually in product management at subscription businesses, Upwork, cook, unity and others. Somewhere along the way, you, you put on the builder hat, you became a guy that shipped a CLI tool, a command line interface tool that pit two agents against each other. And you started putting that inference out on modal in half an hour. Just when did that builder side take over?

[00:01:58] JP: Yeah, it, it is a funny thing. I think I've always been a builder. I think I was a builder as a kid. like it was, it wasn't something that was new that, that, that I started like. A month ago or anything? I, I, early in my career, I was in sales like super early and I hated it just being on calls and sort of picking up the phone and getting rejected and, and doing the whole thing, all the prep work, all the things.

[00:02:22] So it was like really early days of the internet and I was like, I could just put this stuff online. People will find it. It's very naive. People can find it. And then they'll just order it themselves and I could just hang out. and so I set off to sort of teach myself how to code. I built an early website.

Early Website & First Sale

[00:02:39] JP: I didn't know what I was doing, right? Like I, I had it deployed. I was just editing it live in production. this was like 20 years ago. and I ended up like getting a sale before the site was even completed. Like the shopping cart was done with pre-trip, which was awesome. And I was like. Wow. I could just do anything.

[00:02:59] And so like, I, I ended up building that out and, and I had that for a while. and so like that was sort of like the professional builder aspect. We're talking like 2000, I dunno, 5, 6, 7. Like this is, this is a while ago. And then sort of as my career progressed,

Career Path: Startups, PM Work & AI

[00:03:16] JP: in 2015, I started a small startup called Taste, which was like a restaurant and bar recommendation, chat bot way ahead of its time.

[00:03:25] The tech was the hardest part and it's really funny 'cause now I could spin that up. From scratch in like 30 minutes. but at the time it was like months and months of engineering, to, to make that, that work. And that was a lot of fun. That led me into to joining early stage startups as a pm, which eventually led me to Upwork and Cook Unity.

[00:03:43] and, and so although when I joined Upwork, I, it wasn't early stage at the time, but like mostly early stage startup. I ended up transitioning more into the product space because I don't know if you've heard that old adage about PMs, where it's like you either went to school to be a PM or you did a hundred different things and this was the only place that you could sort of touch a bunch of different, different things.

[00:04:04] and so like I was definitely in that second. Second tier and second bucket. so like I just sort of settled into the PM space and I mean, I always approach things from a builder perspective, right? so it was like I was very close to all the engineers that I've ever worked with, close with designers that I've worked with.

[00:04:20] Everything sort of starts from how do we solve this problem. after that point, sort of been consulting and fractional. a lot of early stage stuff, and last year, especially last year, late 2024, getting like I, AI is becoming a much bigger thing. and I, I just sort of got enwrapped in the idea of like, how do I learn more about this?

[00:04:41] How do I do this? and so I sort of jumped into the same mindset I had 20 years ago when I taught myself how to code, which was how do I utilize AI and, and sort of been on that journey since. And, and that sort of led me to wire.

[00:04:54] Jakob Heuser: That's awesome. Like you've got some serious range then. I mean, you're PMing by day. You're doing

[00:05:00] wire by night or by afternoon, depending on

[00:05:02] the day.

[00:05:03] Or on weekends. I guess, what does your week actually look like when you're wearing all those hats? Because that is, that is a lot of stuff being juggled, not only with wire and your growth and your PM work and the fractional work.

[00:05:15] There's also the food rating writing. I've seen you active on threads. there's a product Super Loops as well, like you've, you've got a lot of things in flight.

[00:05:22] JP: Yeah. so agents definitely help. like that, that's definitely like a core part of like how I'm able to do a bunch of things at once. and my client work comes first. That's sort of what I commit to, day in and day out. So, whatever I need to, like contractually do from like an hourly perspective or, or schedule perspective, that sort of is, is my, my starting point.

[00:05:43] Mm-hmm. after that, it's, it's. What needs my attention most? lately it's been wire, isn't always that way. I, I sort of like focus to, to whatever needs my attention. threads is the type of thing I, I honestly do often while I'm like on the subway, like waiting for a trade and, and waiting for it for something else.

[00:06:01] There's like a running joke that people talk about where they like traded their, like Instagram time to using Claude code on their phone. I juggle Claude code on my phone and. Threads. those are my two, like sort of go-tos on, on my phone. but yeah, I, I would never say that I'm great at time management.

[00:06:17] I think AI helps me there 'cause I can just set it off to do things and it does things.

[00:06:22] Jakob Heuser: That's actually kind of cool and it leads, it leads into my next question here, like, let's set aside the pitch for a second. I know as a founder we love to talk about the pitch side of our

[00:06:30] products.

[00:06:31] but you were talking earlier about you how agents have made this process easier. we were talking before today's interview and you were also pointing out that. not only make things easier, but they introduce a lot of friction that people are just not used to dealing with. what's the thing that you see nowadays that makes you mutter at your screen? Or in talking with other people you've seen they're getting frustrated with when dealing with ai?

[00:06:57] JP: Sure.

The Friction People Feel with AI Tools

[00:06:58] JP: the way I use AI and the way I I think about it is this is like effectively a person that I'm working with. I don't mean that in like a anthrop, anthropomorphizing way. I mean that in a, like, this is a, this is a thing that's able to complete a task that I tell it to do. When I like optimize for babysitting it, I'm not really freeing myself up, but if I optimize for it to be able to complete a task to repeatedly compete, complete a task, or to figure out how to complete a task, it allows me to go do other things.

[00:07:33] So I think when, when I see people who get frustrated or have, You know, they haven't quite picked it up as others have. That's the gap that I see there often where they are, not necessarily treating it as like a type of tool where they are delineating work to it. and instead they are trying to use it to do their day-to-day work.

[00:07:55] there's like the people will joke about the idea. Jokes, maybe not the right word, but like people will talk about the idea of taking, copying and pasting things into a chat window and then copying that back into some other sort of application. All of that's really messy. You're not really saving much time.

[00:08:12] but if you sort of trust it and let it go, you'll be surprised that more often things just tend to work.

[00:08:20] Jakob Heuser: It's definitely a, a difference of this is someone that you're working with as opposed to trying to offload, and I think you're, I think you're hitting on something really important there. This idea of, oh, I just want it to do all the work for me, and that's not actually where these tools succeed.

Treat AI Like a New Hire

[00:08:37] JP: Yeah, I think you have to set them up for success, right? Like imagine you were starting, a new role at a company or someone was starting it with you. How would you help that individual get started? You might show them, where they can find document. To read the history of something, you might talk to them about the format of some sort of output you want.

[00:08:59] you might talk to them about where they can go to get coffee. like you kind of wanna set them up in a space where they can be successful. honestly, a lot of that is like where wire comes from. And I, I know you said like, like everyone wants to pitch and that's not me pitching yet. but I think that that's where that is, that is context for how an agent.

[00:09:20] Or how a person can be successful. We just sort of think about it slightly different with a person where we're handing them a guidebook, versus an agent where I'm hoping wire is the tool and primitive that makes sense for them.

What Wire Does: Persistent Context for Agents

[00:09:33] Jakob Heuser: No, that, that makes sense. And wire, just to kind of touch again on wire, this idea of a persistent context across multiple agents, multiple model systems has a lot of value. It's like a wiki page that agents just intuitively understand. And for someone when you're trying to onboard a, an agent, you're trying to say, Hey, I want you to look at this problem.

[00:09:54] I want you to try and fix this problem. there's a lot of value in giving those tools, the ability to go looking. Retrieve that information, catch up on everything and start from a much better, place than just a cold interview or a cold question where now tries to scrape a wiki.

[00:10:13] It can go right to the information needs and get a better start.

[00:10:16] And I think that's Wire's promise if I'm understanding.

[00:10:18] Absolutely.

GitHub for Engineers — But Nothing Else

[00:10:19] Jakob Heuser: I'm gonna bring this back to people. Imagine you started working with a new engineer. What, what is the, the one of the first things that you'd want to do with them? You'd probably give them access to the GitHub account.

[00:10:30] Tell them to familiarize themselves with sort of how things are structured, point them towards documentation. I don't think it's surprising that one of the first major use cases for AI happens to be in the engineering space. GitHub has sort of acted as that like context store for human beings for a very long time, and agents have adapted to that context.

[00:10:51] Or if you think about any other role within an organization or, or an aspect of life. That's not the case. There's no centralized GitHub that exists for it. you know, notion kind of competes in that space. Google Workspace with Google Docs competes in that space. You also have like email histories. You have contact lists, you have other surfaces.

[00:11:14] You have post-it notes and all kinds of things, like it's a mess. Whether that's digital, whether that's physical, whether you're talking about a person or an agent, that that whole piece is really messy. GitHub for engineering makes a lot of sense. that that doesn't exist anywhere else.

[00:11:31] And so just to talk a little bit broader about Wire, then I think with Wire, you're talking not just about the engineering context, but all of these other business documents, all this knowledge management really. Which puts wire closer to this idea of the more modern AI based knowledge management systems, which is, where is this, except it's optimized for the problem solving the builder side.

[00:11:53] If I, if I've got you right.

[00:11:55] JP: it, it can be for builders like the, so like part of the way I've constructed wire is that these containers can be used for lots of different use cases. The, the core premise is there's a database at a vector store. That connects via MCP. There's a whole bunch of software that makes all of it possible, but that is the core of it, right?

[00:12:17] And so like how we think about storing different bits of information, retrieving different bits of information, that's sort of what makes Wire special. the tech is exists, right? So like if you're a builder, right? Can you connect to a wire container and have it. have it, store documents and code and other things and access it.

[00:12:38] Absolutely. That. One of the core use cases that people find themselves, using wire for is maintaining a history of changes similar to like a Git, but at a more granular level, like what was considered, what was chosen, what, what feedback was retained. And keeping that as something that's available to agents that, are, you know, cloud code based, are cursor based, are Codex based, maybe something that they built themselves.

[00:13:06] it sort of plugs into all of those places. But if you're not an engineer, you can of course use it for any sort of data set. I often find myself using it and some, some clients, customers tend to use it for their branding guidelines. They'll use it for their, marketing materials. They'll use it.

[00:13:24] For ICP, management, where, you know, they'll jump on a sales call and be trying to, they'll try to figure out how to best speak to that, that that prospect, that information sits there with their agent, in a wire container.

Wire vs. Knowledge Management Tools

[00:13:39] Jakob Heuser: That's awesome because yeah, there's a lot of products out there that are like that one layer above here's.

[00:13:45] .

[00:13:45] Your AI SDR, here's your AI knowledge management, and you're saying it's just context,

[00:13:52] JP: It is.

[00:13:53] Jakob Heuser: it's better context and making sure it's the right context at the right time on demand.

[00:14:00] JP: Absolutely. like if we think about like agents and how they're, they're structured and, and I don't want to get too technical, but like if we want to think about how they're structured, they are LLM calls that have the ability to make tool calls, which are essentially functions, right? So when we think of what an ISDR is, it is making tool calls that allow it to.

[00:14:23] Research about prospects, potentially send emails to prospects, potentially, uh, maybe update some internal notes, uh, things like that or like the tool calls that exist within this agent. And it's sort of like reasons, like what else do I need to do to sort of get from, you know, one step to the next. Um, and so what's interesting when you look at that sort of agent is you have to find a way to tell it who you are as a company.

[00:14:50] You have to tell it. Who your ICP is, you have to give it. Context around all of those things. What do you sell? How much do you sell it for? What did people say no about previously? Um, that all has to exist there for it to be as successful as an ai SDR. and one of the things you'll find is you go pick up one of these AI SDR tools.

[00:15:13] They'll give you like text boxes to put it in, or they'll maybe give you a connection to HubSpot so it can, to can connect back via their own MCP or what have you. But that information probably should exist in lots of places, right? If you're mirroring that information, between like different tools.

[00:15:29] You suddenly have a problem where now you are managing all of this. It's like, what's the difference between this AI SDR working and you just writing these emails, because you're constantly copying and pasting and changing things. You learn something new about a prospect. Where does that go? it might sit in a HubSpot.

[00:15:45] The, that A ISD has no idea to go back to HubSpot, try and find contradictory information and, and surface that. it gets, it gets tricky. And now. What's really fascinating to me about that is this, your AI SDR has done all of this work. Does it have the ability to take what it's learned as part of this and share that anywhere else in a normal company or sort of in a legacy company, a non-AI company?

[00:16:10] You might find that a sales team will have like a weekly sync with a marketing team, and they will tell them, Hey, like we had these calls, these were objections that we got. We need to update our materials so that they have this information surfaced. Or we need to update our, our, our sales binder or whatever it is.

[00:16:28] or battle cards, right? That link doesn't happen with a SDR 'cause it's like generating the information itself. Maybe it sinks back to a HubSpot, maybe it just lives internally. wire containers are sort of the thing that sticks in between where if the tool supports it, which it should with an MCP,

[00:16:45] It can write back to this container. It's the same container that a marketing team might be using in real time to grab objections that they have received from those, from that outreach to then update their sales materials.

[00:16:59] Jakob Heuser: So then effectively it sits in the middle.

[00:17:03] You've got

[00:17:03] your

[00:17:03] AI SDR that writes to say, a wire container. Instead of it write the AI SDR writing directly to HubSpot. Let HubSpot read from that same data source so that all of your data lives in one reliable source of truth.

[00:17:16] JP: It can be, depending on how you'd like to structure it. I, I still think HubSpot and, and tools like that have a role as like a system of record, right? Where it's like all the things that need to exist do exist here at some point. Is that necessarily the best place for all of your agents to be accessing this information?

[00:17:35] I think one of the interesting places, I keep making comparisons to people, to agents where I think. Agents and people diverges with people. And part of this is because there's no alternative. We don't necessarily entrust them in a like task-based thing. They, we trust them wholeheartedly. We give them access to a tool.

[00:17:55] We, you know, we can sort of segment a little bit. Like you can make certain changes, but not others read certain things, but not others. But this is a, a finite set of permissions that we offer at any given time. And then they have to be revoked at some time in the future. They may not get revoked at some time in the future, but agents don't work that way, right?

[00:18:13] Agents, even long running ones, they spin up, they do the work they're supposed to do, and then once they're done, they stop existing, right? you might have a, a log or, or something saved back to a wire container, but the agent itself no longer exists. So what I find fascinating is like the way we think about authorization and permissions, sharing information with.

[00:18:36] An agent that has to change, right? We don't want to necessarily give agents full permission to HubSpot. We want them to perhaps be scoped to a particular customer or a particular ICP or a particular task. And once that task is done, it should just end. another way of thinking about this is when you're using cloud code.

[00:18:57] Probably I, I would imagine most people listening are probably people who, who use Claude Code or, or some other agents like that. at certain points you'll get alerted and they'll say like, I need your permission to do this thing. You're the human in the loop. we don't have that sort of permission based access for people really.

[00:19:18] In systems of records, they just don't exist. They were never architected that way. and there's probably an argument to maybe that they should not be architected that way. but for agents they can be and should be. And that's sort of where I see a lot of use cases for wire coming about where. I don't need it to have all this information.

[00:19:34] I think it's a security risk. If an agent has all of this information, I need to have some of this information. I needed to have it as long as it's useful. And then when it's no longer useful, it needs to stop existing along with the agent. and so being able to tie those pieces together is, I think where a lot of value for something like Wire and the container system works as opposed to a system of record like HubSpot.

[00:19:56] Mm-hmm.

[00:19:57] Jakob Heuser: That's awesome because yeah, MCP then ends up being like, I mean, it's an API. MCP is just an API, with a lot of really nice to

[00:20:04] have's on top,

[00:20:05] And the API, can

[00:20:06] can become that enforcement mechanism and

MCP, Taskless & the Context-Execution Stack

[00:20:09] Jakob Heuser: I think it's freaking cool. Like why are, Wire is just this awesome tool that helps across the context layer. Obviously the work we're doing at Taskless, we sit at the execution layer . We say hey, these are how we create deterministic rules. You're definitely not allowed to do this-

[00:20:22] JP: absolutely.

[00:20:23] Jakob Heuser: This is, this is by static analysis. Thou have done a wrong thing and the AI fixes it. But the commonality between the two of us is

[00:20:31] MCP, this shared vocabulary , and I'm curious

[00:20:35] where you see those themes between the layers tightening now between context and execution, where you can put instructions in markdown and Claude generally

[00:20:44] follows them of the time. You could, in theory, do static analysis of your skill files to ensure that certain statements are not showing up in your skills, and you could create static analysis of markdown. I'm curious your thought, where are the seams becoming tighter? Where do they need to stay loose so nobody gets locked into a single vendor? What do you see happening in this space?

Single-Vendor Risk & the Multi-Model Future

[00:21:03] JP: Yeah, I mean, I think the single vendor stuff is, is really fascinating. Particularly like we see the, the trend that like Anthropic is on and they're picking up a ton of new enterprise spend and business spend. and they have a great product, right? Like I think Opus is fantastic and Cloud code is a thing that I use.

[00:21:21] All day, every day. so like, I, I, I understand that, but at the same time, like we're building our own agents too. Like everyone's building their, not, maybe not everyone, but like companies are building their own agents. companies are, there is a valid risk assessment to be made where you have a single point of failure with a model provider or a company.

[00:21:41] so when we see like a single vendor, I sit there and see that as opportunity that. Is my entire stack. If I was, if I was running a company, is my entire stack going to be run on anthropic models within the confines of their infrastructure? Is it going to be run in open ai? Is it going to be run in Google?

[00:21:59] Like how does that sit and how does that work? yeah. I just, I don't see that as something that happens long term. I, I think a funny thing that people talk about often is, the idea of like local models versus. Cloud models and, and sort of being able to like distribute that work. What, what decision is a local model making?

[00:22:21] What decision is a cloud model making? How these things, things are working together we're, that conversation is inherently speaking to this idea that we don't necessarily need all the models to be the same. We don't need them to be in the same infrastructure. We don't need them to be under the same umbrella.

[00:22:37] We do need to empower them all. which is where I see a context layer fitting in and sort of where, where I think wire fits in. but I don't see a single, single vendor as like the future of anything. I'll, I'll, I'll give you, I'll take that one step deeper. You, you, you were talking about, you know, markdown documents.

[00:22:55] I don't think of wire as like a replacement to markdown documents. I think they, they solve different purposes, right? A markdown document. Is best used as something that is ingested on every turn, every time an agent makes a decision. clot MD's a really great example of this. It sort of inserts itself.

[00:23:14] It deterministically when it starts searching through a code base will find sub documents and insert itself, but they sort of live there and they exist there. And, and so like you'll often find, you know. more concise Cloud MD documents or agents, MD documents as more effective than ones that are long and filled with a lot of different rules.

[00:23:34] What happens over time is like these rules will get contradictory, they'll become stale. and I think that's one reason that people tend to bind themselves, getting stuck, and so they augment it with skills and other things. And I think skills are reasonable. They make a lot of sense, but that's not the world that like.

[00:23:50] Wire lives in, right? I think these things work well alongside wire, right? It's like, do you want your. GitHub and your code base. Also holding your information around your brand guidelines. Do you want like 150 page PDF sitting in there? Do you want that like versioned in where you're working on a work tree?

[00:24:09] Someone in marketing who now also has access to your GitHub is updating it. But that's not what your agencies, 'cause you have a version like, like all of that's really messy. GitHub was not designed for all of that stuff. wires designed for that. It works, makes perfect sense for a small markdown document that you're inserting at runtime.

[00:24:25] It doesn't make sense for like long living context. Context that needs to live outside of your execution flow. so I think these things are very complimentary.

[00:24:35] Jakob Heuser: I'm really

[00:24:36] curious about your other thoughts then.

Will Claude Just Absorb This?

[00:24:38] Jakob Heuser: I don't think Claude and other tools are very far away from being able to share memory across sessions or across teams.

[00:24:43] JP: Great.

[00:24:44] Jakob Heuser: Other than the fact that it writes to my home directory every time I

[00:24:46] tell it not to, but that's a separate concern. Boris,

[00:24:49] if you ever listen to this podcast, please, when I say don't write to the memory directory, don't write to the memory directory. but memory exists on tools

[00:24:57] JP: Yeah.

[00:24:57] Jakob Heuser: and I don't think it's that far from memory being something that Claude begins sharing between sessions and on enterprise plans between teams.

[00:25:06] And I guess I'm curious, where does wire go from there? Because I think it's bigger than just the context, but that's obviously where you started.

[00:25:14] JP: Right. And, and so like, I, I think a couple of things I, I fully expect that to happen and I think it does, right? Like there is like a cloud code memory tool that's built in saves to mark down files. I think there's value there. I'm not saying that there's not, I think that is. One particular use case.

[00:25:32] But the thing is that Claude Code and, and Anthropic and Open ai, and all of them who have their own like tools and memory don't exactly have the greatest incentive to open that up across multiple providers, right? the memory that I create in Claude Code ends up being a moat. For more usage in cloud code.

[00:25:49] It's not necessarily designed in a way that, you know, I can open up a session in Codex and obviously this is purely in an engineering context, that I can open up a session in Codex and, you know, ingest that same memory and use that and pick up like I've never laughed. I imagined whoever you know at some point is sitting in the, in second or third will try to do these things.

[00:26:09] but whoever's in first and whoever owns that market share, is not going to be doing those things. they have no incentive to. And you, we, I don't think we, I can't remember a time ever truly seeing that. we, even if we look at like Gmail, which is like number one, and we go back to that as sort of a thing, it kind of works on my iPhone, but it doesn't work the same way as it works on an Android.

[00:26:30] Right, and so like we sort of see these gaps in, in all time with software and I would expect to see those going forward here as well. So yeah, like right to markdown, I think it makes complete sense. For that specific use case. I don't really see those as competitive products. Now, does that like eat away for some of the use cases?

[00:26:52] that's not the only thing, right? I, I think one of the nice things about Wire is that it's designed to ingest any kind of context at any time from any source. just yesterday, well, depending on when this comes out, this, this might be really old news. Built in support for our, for using N eight N workflows as a source of information that can flow into a wire container.

[00:27:15] Now, you could build something within cloud, go to go and do that and set that agent to run long term. And there's a lot of like orchestration that you then have to set up to do that. That's not the case with Wire. It's, it's like one click, there's a, there's a plugin. You, you set it up and it just like ingests that information.

[00:27:35] You'll find that like some companies, will, like Claude, even I keep using Claude because I use Claude. they ship with connectors to like Google Drive. and you'll see that as like a, Hey, this is a universal connector. I can keep my files there. And yeah, I think on some level that makes a lot of sense.

[00:27:50] If you have documents in Google Drive, you should want to be able to access them. I still don't see that as a real competitive product because every time you go and make a search through Google Drive. It has to process that information from scratch. You have this million token context window. You're now going to blow 200,000 of that so that it can search through a bunch of, like doc files, or, or what have you, or, or sheets files to find the very specific piece of information that you needed to find.

[00:28:20] Or you better yet you go to Google Tribe, find it yourself, then tell it specifically which file that it has to go to. I, I, I just, I don't see these, these are great tools. But these are tools that existed for people prior to agents. And so wire is built for agents, which means when you add a document, add a piece of information, and for what it's worth, we support two-way sync, meaning your agent can write back to the container, that information's processed once and then it's available.

[00:28:49] You're not burning through. All of your context window to go and find a document, process it, convert it. It's not writing a Python script so it can like understand a document that's sitting in your Google Drive. All that stuff's happened already, right? It, it's, it's very easy for an agent to just quickly pick it up in a token efficient way.

[00:29:07] and there's actually one thing that I, I think, I don't see this a ton elsewhere. I think some systems of records do this, but in very specific ways. One of the things that I love about what Wire does in the container is. It builds two graphs. Right. And, and I think a lot of like GitHub projects that you'll find that talk about memory and things like that will speak to one of these graphs, which is a topical graph, which is like, what are the pieces of information?

[00:29:37] If I was to like build you a blueprint of what exists in terms of information, what, what's there and, and that sort of exists. And, and, and obviously an agent can go and query against that and get filter, sort, search, whatever. Um. It's useful in, in my experience, in my testing and in the evals. It reduces token spend, reduces turns.

Wire's Epistemic Graph: Tracking Knowledge Over Time

[00:29:59] JP: But what I'm really excited about is, is a epistemic graph that we created, and what that's designed to do is. To take any individual claim that gets inputted within a wire container and try to determine if, how that information relates to other information that already exists there. Does this corroborate a different piece of information, a a different claim?

[00:30:22] Does this contradict a different claim? Has this superseded a different claim? like, like really fascinating, right? Like so, so Tim Cook is CEO of Apple. Or was CEO of apple, or I guess he's transitioning out right now. If that lived as a claim within a wire container, when that changes, we would not have to go there and find it, delete it, make changes.

[00:30:45] we would instead just add that now there isn't new CEO of apple, and that information would then get tagged as superseding the previous information where Tim Cook is the CEO of Apple, which means that. wire is not making determinations of whether information is what you should use, but it's surfacing that.

[00:31:06] So if you, if your agent was searching for the CEO of Apple, it will pull up both pieces of information. One will specifically be tagged as superseding the other, and your agent can then make that choice and determination to be able to say that yes. This new person is the CEO and it's no longer Tim Cook.

[00:31:25] and that doesn't really exist. You pull a bunch of markdown documents together and it's gonna search through that. It'll look for CEO, but it doesn't know exactly, like, Hey, when was this written? Was this written sooner? Was this written later? Was there a date attached? Is there, how does this information directly relate?

[00:31:42] It's, it's doing all this by itself. It's figuring it out and. I don't know what it'll give you as an output, but in wire it's very, it's labeled right. It it, that information is there for your agent to be able to make a proper determination on.

[00:31:57] Jakob Heuser: Which means it actually tracks the evolution of knowledge, which is who uses a lot of personal knowledge management systems, like I have an obsidian with a Z Costin system running. of tracking how the knowledge is evolving is so vital to that context, which is obviously what Wire specializes in.

[00:32:14] JP: Absolutely. I, I think a lot of, a lot of people, I think a lot of like interesting, products that people create that, that live in this space are a little too. specific about how that information should be used. Like they will expire an archive and delete other information that it, like deems contradict.

[00:32:33] Well, this is newer. It's contradicting this, this should be deleted. This should never be surfaced. And I, I can understand if you have. A full view of information and you are explicitly adding this information in that you might want that as an output, but at scale, how can you be sure that's the thing you want?

[00:32:52] Right? Where I have a container that's long running, it is just pulling in, let's say sales leads at the back end of an N eight N pipeline. Two people from the same company wind up in this container. If it was just writing to a markdown document, it just sort of exists there and it's sort of like, oh, well it's on this line.

[00:33:10] I guess that means it came later. Is this how it's deter? Like your, it's up to your agent to go figure all that out, right? And, and so in a wire container, that information is labeled, it is accessible. And, and I think it makes all the difference for your agent to be able to make the determination for what is the task that it's doing.

[00:33:29] You know, if it is to send outreach emails, yeah, you do want the newest person, if it is to update someone, get them up to speed on how a deal is progressing. You want them to know that this person was there before, but they're no longer the person that you should be talking to now. so like agent will have a different use case depending on what the query is, and.

[00:33:50] A proper epistemic graph will provide that information so that that agent can make the decision.

[00:33:56] Jakob Heuser: That that actually kinda segues well, like I don't think in 30 days epistemic graphs are gonna be the new hotness that everyone's talking about on threads. this podcast does drop about a month from when we're recording right now. And so I want to hear from you as someone that's in this space, deeply plugged in, what's gonna be the topic we're gonna be talking about in 30 days?

[00:34:18] I mean, other than open AI and Anthropic having another Twitter slap

[00:34:22] JP: Yeah. I think one of the core realizations that we're going to find, and I think if this drops in 30 days, we probably would've shipped something for this too, is retrieval, And context are so critical to workflows and agents performing successfully, that we'll be able to do a lot more with great context and a lower quality, more efficient, smaller agent or smaller model powering an agent.

[00:34:55] and I think that'll be the talk, like how, how good does context need to be that we can run this on edge? Without requiring me to spin up like a, like an opus level frontier model to actually do the task that I, that I required to do. I think we're seeing,

[00:35:13] Jakob Heuser: I see, I think I see some of that already. Like where you

[00:35:16] have.

[00:35:17] I've seen, people talking about how they use or opus as an orchestrator.

[00:35:20] JP: Right.

[00:35:21] Jakob Heuser: Then when it comes time to do the work, you don't ask Opus to

[00:35:25] JP: to do the work.

[00:35:25] Jakob Heuser: Because if Opus can give you well-defined plans for the work, can hand that out to a smaller thing through ollama or something like that and let it just it cook.

[00:35:35] JP: Absolutely, and, and I think we'll see that even more as we go forward, where frontier models will still be amazing and incredible and useful for a lot of engineering heavy tasks. A lot of fact finding heavy tasks, but. Models that are trained to retrieve knowledge, will end up being models that we hope to use on a day-to-day basis because they'll accomplish the tasks that we need them to.

[00:36:04] They'll be the AI SDR, right? And then they'll have this context and be able to draft this email and send it out or get it ready for you to review it, to send it out. And you won't need to be spending, you know, $15 per million output tokens.

[00:36:20] Jakob Heuser: That makes a lot of sense. What's your daily drivers then? So right now I know you've got Cloud, cloud code open

[00:36:27] JP: Yeah.

[00:36:27] Jakob Heuser: I know you've talked about Cursor, you've talked about up to Code Axon, flow Ninja. You've been through a lot of tools.

[00:36:33] JP: Yeah.

JP's Daily Stack (No VS Code)

[00:36:33] Jakob Heuser: What are you using nowadays and what have you actually set aside?

[00:36:37] Like what, what's a, where you're like, you know, this actually isn't working and I've set this aside for me.

[00:36:42] JP: I think it's really funny, like if you, if you sit here and like if a person like looked at. My computer, I'll be like, what do I have open? Like literally, what do I have open right now? I have Cloud Open. actually I have Cloud Open in two ways. I've clogged the app open and Claude in Terminal and Warp open.

[00:36:59] I have a calendar app open. I have Chrome Open, I have Slack open, and have granola open. That's the other app that I have open. Everything else closed. iMessage, I guess is technically open too. This, this is my like, daily stack. I, I like, honestly, wish I had something like super cool to be like, oh, I use this weird app that no one's ever heard of.

[00:37:20] no. Just very simple basic works.

[00:37:24] Jakob Heuser: some, there's something cool missing from that stack vs. Code isn't in that list.

[00:37:29] JP: Exactly!

[00:37:30] Jakob Heuser: I, I think that's actually really, really important and kind of cool. Like you don't, you don't need to, you operate through diffs.

[00:37:36] JP: Yeah, I, I can't remember the last time I specifically opened up an IDE to make a change. I viewed stuff, of course, like, you know, the viewing in most of these applications is pretty rough, and github.com is not a great viewing experience either. So like, I've definitely opened up an IDE to like view a change, but usually not open.

[00:37:58] Jakob Heuser: That's actually really, really cool. Like out, out of all the things in the list. I didn't expect vs. Code to be the one. Like,

[00:38:06] yeah I just set.

[00:38:07] Set my

[00:38:07] IDE aside I really don't

[00:38:08] use it anymore.

[00:38:09] JP: Yeah, I mean, like, Claude code through terminal usually does a, a fairly good job. things change so fast. like I, I think some of my diffs end up being, you know, like 10,000 lines removed, 11,000 added. I'm not overly concerned about. I, I'll say this, right, there are some rules set up, right?

[00:38:30] Like it, like it's not just, you know, sort of writing whatever it wants, whenever it wants. There, there are proper rules set up. There are tests that it has to write like, like that workflow exists. but it's not like I, I'm not sitting there saying like, oh, you know, this, this function got written twice.

[00:38:46] Like, I'm not sweating over that because we'll probably not be there in a week, because it would've been written over again by something different because we've changed something, we've added something, remove something, whatever that is.

Tech Debt as a Linear Ticket

[00:38:57] Jakob Heuser: That's actually fascinating. I mean, we haven't really talked much about tech debt on, on this

[00:39:03] podcast.

[00:39:04] And I think that. that comes home to roost, but I think it comes home to roost once you're an established company, like once you have enough paying users that the tech debt begins to slow you down hopefully by then Cloud code will have a, have a tech debt tool and we can just one click erase all our technical debt back to building.

[00:39:21] JP: know what's actually maybe a better question than like what I have open, or what, what applications I use is. What CPS is my agent connected to? Right? And I think like this, this bears a lot, right? So like, obviously I have wire containers, I have multiple wire containers connected. I also, one of the tools that I keep connected is linear.

[00:39:43] so what ends up happening is oftentimes when you know it's doing something, it comes across something. There's tech debt, there's something that should be handled, something that should be solved. I'll have it. I'll have cloud code just generate a. a ticket for it, save it into linear. Great. but what ends up happening is like overnight when I'm not using my usage, I will assign it 3, 4, 5, 6 tickets to just have ready by the morning.

[00:40:09] And a lot of the sort of background like, Hey, we need to make sure we get right tests for this particular thing, or I need to clean up something that's sort of like hanging an old workflow that's, that was sort of like half built but never completed. Yeah, these are assigned tasks overnight. I review them in the morning.

[00:40:29] sometimes it's a Sunday morning and I just ship it. and so like I, I, I don't necessarily think of tech debt as something that I sit there and do it when it's identified, it winds up in a linear ticket and then I just assign it as like additional tasks at some point.

[00:40:43] Jakob Heuser: That's awesome, and you get to use up your usage at night when you weren't using it

[00:40:48] JP: Exactly. I have a five hour window. I'm hopefully sleeping for five hours.

[00:40:51] Jakob Heuser: I hope so. if, if,

One Sentence Takeaway

[00:40:58] Jakob Heuser: if everyone's listening today could take one sentence away from everything you've talked about, what do you want that one sentence to

[00:40:59] JP: Context rules everything around me.

[00:41:02] Jakob Heuser: I love it. You have that on a shirt,

[00:41:04] JP: I don't yet, but I probably should. I, I, I think I will before this podcast comes out.

[00:41:10] Jakob Heuser: Awesome. Awesome. I love it. everybody, JP . Absolutely wonderful talking with you. Everyone should go check out wire use wire.io. I'm Jakob, I'm CTO at Taskless. This has been The Task at Hand. We're doing a podcast once a month. Always stay tuned. We're gonna have all kinds of content. We're gonna talk to builders and makers about the stuff they put into the world. Thanks a lot, everybody.