Speaker A: I don't know why I just suddenly felt strange on my—

Speaker B: Well, it's because it's recording.

Speaker A: Hi everyone, you're listening to Scaling DevTools. I'm joined today by Dennis from Unblocked.

Speaker B: Hi, thanks so much for having me. Great to catch up in person.

Speaker A: Yeah, so it's the second time that you've been on, but yeah, this time we have a grand piano in our background.

Speaker B: We're looking for a place to sit and we could only find one beside a grand piano.

Speaker A: That's not too bad. So we're at Lead Dev. So I should thank you, Lead Dev, for having me, allowing me to do this. How have you been finding the conference so far?

Speaker B: It's phenomenal. I think it's one of my favorite conferences this year for sure. We've been to a couple that I really enjoyed, but the level of talks, the organization of the event, the people, it's fantastic. I think it's just like— yeah, it's definitely a highlight.

Speaker A: And with the only addition next time that you're gonna bolt on some surfing, hopefully.

Speaker B: Yes, we were chatting about surfing. I'd love to maybe add a day on either end of the trip to see if I can get some surfing in, in the UK for sure.

Speaker A: Would love that. Which is very good. Everyone listening should go surf in the UK. For sure.

Speaker B: Bring your wetsuit.

Speaker A: Bring your wetsuit. Bring a thick one. Yeah. Yeah. So could you tell us, I know you've been on kind of a journey since we last spoke. Could you tell us a bit about Unblocked and what what you're working on.

Speaker B: Yeah, I mean, it's been a crazy, I don't know, maybe 6 to 8 months, 10 months or so. Um, Unblocked originally started as a platform— well, let's take a step even way further back. Up until 6 to 8 months ago, people were the, were the beings that actually went and wrote code, right? And so in order for them to be able to write code in their business, in their organizations, they needed a bunch of context. And so Unblocked has always been this context layer that had interfaces where people could ask questions to get answers. 8 months ago, 6 months ago, thereabouts, towards the end of late last year, I would say, we saw that it was increasingly these agentic workflows that were becoming quite popular, and it was increasingly agents writing code instead of people. And so people were helping that process, but agents needed the same level of context that people had. And didn't have it available to them. And so we've been focusing on making agents just far, far better with servicing that context layer and making it available to them. So what we're seeing is agents are getting their jobs, their tasks done between 40 to 60% faster with Unblocked, and we're seeing significant token savings anywhere between 30 to 50% less token usage for a given task. So Yeah, that's a little bit what we've been thinking about.

Speaker A: Okay, and like with, I guess context is like a big topic. Yeah. And yeah, yeah, what are the parts that are like, are there any interesting things that you've like discovered about like what really matters in terms of context?

Speaker B: For sure, I mean, I like this expression where it's old is new again, right? So if you think about your first day at a company, you had no idea how the codebase worked or why it worked the way it did or how decisions get made or any of that information, right? And so in some capacity, you had to build all this context as a consequence of digging through information or interrupting your coworkers, opening pull requests, getting feedback on your code, participating in postmortems for outages hopefully you didn't cause, but maybe you did, right? These are all the lessons and the context that you built as a consequence. And so Over time, you've gotten to be really good at your job because you have all that context. So all of that information is what is missing when an agent gets spawned in a terminal. Every time you spin a new agent up, it's like you on day one at that company. It has no idea about how these things work. And so what it does is struggle. It does a bunch of unnecessary tool calling or grepping through the source code. It has a whole bunch of limitations in the way that it's trying to get the information it needs in order to get its job done, right? So we've been like, that's kind of the mental model that I think about, which is like, you know, you open a new Claude code terminal. What does it know about your organization? Nothing. It has to learn it all again, right? And so like, we see people go through these phases of like, you know, spoon feeding it context to get them up to speed, basically.

Speaker A: Yeah, and I guess like the challenge that I saw working with context was like we would try to— trying to decide like what— because we were building something that like used context and it had to know about their business and stuff like that. It would be like— so things would kind of contradict each other sometimes and like it was like trying to do like help keep memory of like, okay, this thing happened, but maybe that information becomes out of date. And like, so do you keep it around? But then you decide in like which things are still relevant and like how this kind of things are.

Speaker B: You know, it's been super interesting over the last probably year or so seeing how organizations adopt AI, agented coding specifically, AI for their agented coding scenarios. What we've found is basically there's kind of a maturity model. Right? Because different organizations do it at different rates. And even within the organizations, different people do it at different rates. So in the most nascent version of that maturity model, I think it largely stems from people using Obsidian or collecting a bunch of files that they have locally that helps those agents understand how they should perform.

Speaker A: Yeah.

Speaker B: Right? I think that's everyone's like, yeah, I can totally recognize that. And then at some point they sit down and work with someone else., and they're like, well, this is how I think about context, this is how you think about it, so now we have to merge these two things together. So you're hand-curating this context, and you check that in probably, and it is now basically a static data source, right? It's a static chunk of information that you have to now manage and maintain. I think everyone's had some experience with Confluence at some point, or any kind of wiki or document, moments after you're finished writing the document, it's obsolete, right? And so then you have to find someone to manage that. And at a sufficiently scaled organization, finding someone who can apply the judgment that makes sense for their entire organization is actually really hard. So this idea of hand-curating context gets you a level up in terms of the agent's ability to do their work, but you kind of hit a ceiling. So people are like, well, I have this static data source. The most natural thing to do is like now surface the context through MCP. And so like the next kind of stage is like, I'm going to expose all of my line of business systems to these MCP servers. Well, that again provides you a level up. But what ends up happening is that the agents don't know how to think about data across a whole bunch of different services. Right? There's this term that we've, that we've kind of like really found a great way to describe it. It's called satisfaction of search. Its origin was actually from radiology, and it's basically the phenomenon where radiologists would unconsciously dismiss any potential other anomalies once they found the first one, right? So like, think about it as a patient, you would not want your radiologist being like, oh, that's it, we're done, let's move on. You want an exhaustive search through all the imaging to make sure that like they've identified everything that's required for them to come to a complete conclusion. Agents do the exact same thing. They'll hit the Notion MCP, they'll hit the Slack MCP, they find something that looks right, and they carry on their merry way and don't go to the exercise of making sure that it's actually fully correct and finding information that actually might change that conclusion, right? So like MCP is a great enabler to help you get rid of this kind of content rot that happens when you hand curate this stuff. But again, it gives you— it doesn't give you the exhaustive truth. There isn't like a fully reconciled view of the thing that the agent is looking for because it's only finding in one data source. So then the next final stage is basically building a content layer, right? Which is— or a context layer, I should say. And that's basically the idea of pulling all that information together, creating a knowledge graph, and then making it accessible for agents to call into and get like answers right away, right? They're completely reconciled across all these different data sources across your entire team.

Speaker A: Yeah, this seems really valuable because I remember some of the challenges we had was like time-based information where it's like gets superseded out of time. And then also stuff like, I always thought this was an interesting one. I asked it like, "Tell me Jack Bridger's secrets," and it would just like use a search— this is not context necessarily— but like use a search tool to like search Jack Bridger's secrets. And it was like, "Oh, I can't find anything." And it's like, well, if you were like a human tasked with that, you'd be like, "Hmm, I don't know, like how could I uncover this guy's secrets? Maybe I'll like listen to the podcast transcripts and like find—" but it would just do like kind of dumb stuff. Superficial, yeah. Terms or something. And like, is this the kind of things that you guys are sort of—

Speaker B: Yeah, the agents, I mean, the agents are like phenomenal at writing code, right? But they don't understand necessarily a lot of the intent behind the code that they write. And so it's often misconstrued, right? So, you know, these, I think of them as like experts strangers, right? This other kind of mental model that I have is imagine you have two people who work in— who have 10 years of experience. One of them has been at your company for 10 years. The other person is new to the company. If you ask both of them to fix a bug or implement a feature or make some architectural decision, who do you think is going to be better at it? Why do you think that's the case? I mean, they have the same skill set, right? But it's the one with the context, right? Context is— intelligence is not the gap. It's context that's missing. So that's the thing that we've been kind of working on and making available to both people and increasingly their tools to write this code. So the wins are very, very quantifiable. Yeah. Have you been able to—

Speaker A: I guess how you get the stuff that exists in people's heads that isn't even written down, is it about? That's harder.

Speaker B: Yeah, for sure. Right. Because it's like, ultimately, if it hasn't been transcribed in some way, shape, or form, you know, we don't yet have, you know, brain to computer. Yeah, brain to computer. I think the Neuralink folks are kind of working on that or whatever, but we haven't, we haven't used that at Unblocked yet, I would say. But it is, yeah, it is limited to, you know, Even with things that are written down, inevitably I think one of the hard parts about building this context layer is being able to find the truth when you have data sources that conflict with each other. Yeah. Right? So there's all sorts of fun problems when you even have all the data available at your fingertips.

Speaker A: Yeah. What do you think have been the biggest, like typically like the biggest wins for teams that use you guys versus like DIY, like? Failed at managing context themselves.

Speaker B: Yeah, I just like our favorite kind of exploratory phone call with a potential customer is someone who's tried to do this themselves. Yeah. Because they're like, yeah, this sucks. Yeah, it's super, super hard to do it for any number of reasons, right? I had someone ask me yesterday at the conference, like, what's the hardest kind of like technical problem that you've had to solve? And there's just like such a huge list of things to go and do, but I'll give you one, you know, one kind of fun one, right? A person's identity— a person can have multiple identities in each system, right? You have an identity in GitHub, you have an identity in Slack, you have an identity in Notion, you have an identity in Confluence, whatever it might be. All those systems of record have a different kind of understanding of who you are. Yeah, but in order to like be able to reason across all those things with respect to what access you have, there needs to be a unifying identity. So like, how do you create an experience such that someone who, you know, what's the business problem they're trying to solve? Well, not everyone has access to all the information. So an organization might choose to enforce a permissions-aware version of unblocked. Where they're like, you can't ask the question, how much money does my boss make? You can ask the question, but Unblocked can't give you the answer because you don't have access to the Excel or Google Sheet or whatever it might be that contains that information. If you have access to it, then Unblocked will answer it to you. But in order for us to actually solve that problem, we have to be able to resolve Jack to a series of identities across a bunch of different systems and at runtime be able to deduce what he has access So there's all sorts of these kind of hard problems to go and solve. But when you do them well, it's completely transparent to the user. They just get incredible answers very, very quickly. And it's completely transparent to them. It's my favorite type of software, doing really, really hard things in a way that we have no appreciation for.

Speaker A: Yeah. That's amazing. Just check your— OK, we're about 50, 2:50. Yeah, we're good. OK. Yeah, that's really interesting. One of the challenges I remember was kind of knowing if it's an improvement, I guess, as well. Like, when you make changes to how the context works, I guess this is probably a broader problem, but— What do you mean? Like, knowing if you make changes to how the context is managed, to know if that's better than what you were doing before. Doing before, or is it a part of what you do as well?

Speaker B: For sure. I mean, I think a lot of people— I guess it depends on how the context is managed. One of the approaches that we've taken is that you shouldn't have to change your workflow. And it's really hard to change an organization's workflow for this software to work, right? I've never met a team that feels like their documentation is like, well organized and up to date. Yeah. It just doesn't exist, right? Everyone feels like, oh, you know, this maybe should be cleaned up, or that's out of date, or whatever it might be. And so it's on us to be able to connect to those data sources no matter how people work and make sure that it is palatable for the processing that needs to get done, right, to provide these high-quality answers. Yeah. Yeah, so I think it's a very unreasonable thing to ask an organization to go through all their docs and remove all the stuff that's obsolete. It's just not going to happen. Or like, you know, go to Slack and delete all the Slack messages that are no longer relevant. You know what I mean? Like, people know what's— Interesting. Yeah, yeah. So like, we actually just say, don't change anything. We will connect those different data sources. We will clean and enrich the data completely transparently. So when we build this knowledge graph for the organization, it is already— it's literally uplevel the quality of the information or the travel knowledge that exists in the company. Yeah, it's—

Speaker A: there's some— I don't know if you ever tried one of these like email, like AI email management tools. Like Fixer or something? Yeah. But it's like they kind of— I feel like the way they work is they'll like completely change the structure of your email. So it's like everything is like categorized and stuff. And it's like kind of— it's really— my friend works there. I think they're doing really well and it's really good. But it's like it is changing the— it's fundamentally changing your email. Inbox, the data within it and stuff, and like, because I guess that's how they manage their context, because it's just too big to like be searching Gmail constantly without having all of those like filters and stuff on it. I guess you're saying like it's like not modifying like the underlying data.

Speaker B: Yeah, I think like if you're gonna change the way that someone works there needs to be a very obvious and near-term outcome of that change. Yeah. Right? Which is like, I'm gonna rearrange your email, and all of a sudden when you engage with your email, there's something delightful that immediately happens to you. Yeah. But changing how you have a conversation in Slack, or forcing you to go and update all your historical Confluence documentation, or any of that kind of stuff, you know, something that you might see a benefit of should that happen to be relevant to a conversation or a question that the AI uses moments, minutes, hours, days, months later. It's really hard to connect those two things. So like you have to do a whole bunch of upfront work for a payoff down the line. I don't think that's going to work. I strongly don't believe that's going to work. So we're just like, don't change anything. Yeah, we'll do all the heavy lifting. And what happens to you is you ask questions, your agent, like, yeah, you just get the answers that you need completely opaquely, right? Super cool.

Speaker A: Yeah. Okay, and I think probably final question. Yeah. What are you excited about at the moment, Dennis? Oh, besides surfing.

Speaker B: I was going to say, I was like, we spent the first hour, half hour chatting about surfing, so I'm like, I have a little bit of like, I think I've been primed to think about— Surfing in Cornwall. Yeah, yeah, exactly, surfing in Cornwall. I love to go surfing. What am I most excited about? We've been talking about context for years, and so every idea has its time. Yeah. And it's really, it's cool to see people latch on and see that like, oh, this actually really, this is a big, like, people like, I have that problem and this is a big lift for us.

Speaker A: It's nice to see that for sure. It's amazing. Yeah. And where can people go to learn more?

Speaker B: getunblocked.com is where you can learn more about it, or you can follow me. I have the longest last name, Greek last name you'd ever think of, but it's Dennis Pilarinos on X. It's not that long. Oh yeah, well, if you type in Dennis Unblocked or Dennis Get Unblocked, you'll find me for sure.

Speaker A: Brilliant. Well, thanks so much, Dennis.

Speaker B: Awesome.

Speaker A: Thanks for having me. Thanks everyone for coming.

Speaker B: Good to see you again, yeah.

Speaker A: Good to see you. Are you going to play the piano to see yourself? You first.