How Intercom 2x’d their engineering velocity in 9 months with Claude Code | Brian Scanlan

[00:00] Suddenly you started realizing that you have to think bigger about things or that your imagination is now the barrier, not the tool. How is this not happening in your organization? Like literally the physical limits of my ability to type code are unlocked by AI. Today we are seeing twice the number of throughput as we did compared to nine months ago on our engineering team. Now it's like why can't it be 10x? This is a little bit more of what my instinct tells me is possible, which is if you go all in, if you prepare your team, if you prepare your code base, I think your overall [00:30] product quality is going to go up. I think your overall developer experience [00:32] experiences going up. There's just so many good things that come out of using these tools and using them correctly. Backlog zero is a realistic thing for teams to be able to go after. All the things that you wish you ever wanted to do, it's now just achievable. I often advise a lot of CTOs and VPs of engineering when figuring out how to get their engineering team AI pilled, say, everything you hate about the code base, go spend a month fixing and see how fast we can speed run that. That's going to feel really good. I've been having the most amount of fun in my career over [01:02] wants. Welcome back to How I AI. I'm Claire Vaux, product leader and AI obsessive here on a mission to help you build better with these new tools. [01:14] Today, I am showing how Intercom 2xed the number of PRs that their R&D department is shipping in just a few months. Brian Scanlon is a senior principal engineer at Intercom, and he is going to show us truly [01:29] all of their secrets to getting a large product and engineering organization

[01:33] cooking on Claude Code. [01:35] Let's get to it. [01:37] This episode is brought to you by Celigo. Every company today wants AI to improve how work gets done. The fastest way is building it directly into everyday business processes, automating employee onboarding, keeping customer data accurate, managing orders and inventory, or resolving finance and operations issues. When AI lives inside the flow of work, it can update records, trigger approvals, route work, and kick off the next step across systems. That's how teams [02:07] and deliver measurable results. [02:08] Soligo makes this possible. And now, with Soligo Aura, it's never been easier. [02:14] Soligo Aura gives you access to the entire platform through natural language, connecting your systems and turning intent into action. [02:22] All of it under your control. [02:24] Companies like Databricks, PayPal, and Olipop rely on Soligo to run critical business operations at scale. [02:31] Ready to operationalize AI? Visit celigo.com slash howiai. That's C-E-L-I-G-O dot com slash howiai. [02:44] Brian, welcome to How I AI. Why I am so thrilled that you agreed to join the podcast is I think [02:53] Intercom has done it, which is... [02:55] you all have met the moment in sort of two ways. One, [03:00] Clearly, I'm at the moment from a product perspective where one of the first companies that had, sorry, I don't want to say legacy business, but had a going concern business.

[03:09] that saw AI coming and really transformed how your product worked for customers. And I'm a happy Finn customer. They did not tell me to say that. And then second, what we're going to talk about is the team at the moment in terms of really understanding... [03:22] AI was going to change how, in particular, product engineering and design orgs and engineering organizations were going to work. [03:29] And you just went full speed. [03:32] at changing how the team works. [03:35] What drove sort of the urgency around meeting the moment? How did that come to be? Was it a single person? Was it everybody? What was your experience? [03:44] I think in some ways it's been... [03:47] the easiest place to be driving out the adoption of AI in engineering and product, because we've focused the company so much on our folks and product on adopting AI and being AI first and how we think about the product, future customer support and all that. And [04:05] We also had very clear expectations. We've seen what's possible in the product space, [04:11] And it's just very clear and obvious to us as like connoisseurs of AI. It's like, this is clearly going to be huge in engineering and product and building. And honestly, there's been a lot of impatience for like, why isn't this happening today? You know, if we go back a few years and cursor is picking up a bit of business and the models are getting better, but it still wasn't transformative. It still wasn't like the whole business was changed and we're seeing vast amounts of extra productivity.

[04:41] But it's still... [04:42] felt like we needed to have some sort of breakthrough moment or something needed to big had to happen for us to get to the kind of huge velocity wins that I think now we're starting to achieve. That said, we still want more, you know, we're proud of where we're at, but we're [04:58] we're not content with what we've achieved so far. I feel like every three months I have a [05:04] A breakthrough moment. [05:06] And in fact, I feel like Opus 4/6 [05:10] I don't know, something just like really inflected in... [05:13] what was possible when that particular model came out. Now I think the GPT-54 [05:19] Models are also exceptional. And so it's something about that one moment with models [05:24] that really inflected my own personal use of AI and engineering. Did you all see the same sort of inflection around that model model? [05:31] Point. [05:32] Totally. I think you can... [05:33] go back to, it was like November, December last year. And suddenly you started realizing that you're, you have to think bigger about things or, or that your imagination is now the barrier, not the tool. You're spending less time massaging the tool to get it to the right place. And it's less about autocompletion and more about just literally giving us your ideas and seeing what happens. I think the Christmas break happened as well. I remember we, we pretty much decided before Christmas, like, Hey, we're going to go all in and cloud code. Cause up to that point, [06:03] a bit of cursor here and there and augmented different tools and the christmas break really helped like we uh but i just saw everybody go wild on twitter x you know that uh people were uh talking about how this was like they were getting so much done and they were building all these things i just come back to work after christmas break going like okay everything's changed like we we knew that there was something here and that we're starting to see the signs of us but now the whole world is convinced or at least all of the influencers on twitter and i like that that

[06:33] - I'm actually kind of convinced that [06:35] companies should increase their PTO and parental leave policies. Because everybody I know right now in tech that is quote unquote taking time off, [06:44] goes on their vacation and pops open Claude Code and comes back like 10 times more skilled than they were before their time off. And so if anybody wants a little minor hack to AI literacy in your org, give people time off to hack. And they will come back with more information than you expected. OK, I think we're going to skip to the punchline, which I love, which is, [07:06] We're going to see how AI has actually changed how you all ship. [07:11] So can you just show us a little bit of how this has changed inside the org? And I think you all are measuring a lot of this. Yeah. So I think we've been diligent as... [07:20] you know, product owners inside of Intercom that we've been trying to [07:25] get feedback from people and see how they're using the tools and really like just [07:30] doing everything that we would normally do with a regular product. And so we've spent a lot of time hooking up cloud code with telemetry, both into things like Honeycomb and data also going into Snowflake, where we have our data warehouse. And we also store session data in S3. And we mine this stuff for useful insights. But one of the main things that we used to drive adoption [08:00] our CTO Dara setting a goal of us 2xing, like doubling the throughput of R&D. And we use pull requests as a crude, simple measure. But, you know, you can argue back and forth about what's a good measure, what's a bad measure, and whether measuring anything's appropriate or whatever. But I think it's reasonable to just have the expectation that if you can get a lot more done, and it's so fast and fun, then why isn't everyone just shipping more stuff?

[08:30] adopted and that they're being used well. And of course, we don't tolerate lowering quality. And we're high-trust environments. So we don't expect people not to game these stats or whatever. But our metrics-- and what I'm showing on the screen here-- is it's a classic number goes up kind of thing. Where we started tracking this back [08:50] How many... [08:51] PRs and what percentage of them were generated by either Claude or Cursor or whatever. And, yeah, since our major investment in Claude Code, the platform, I'm going all in on it and really pushing out, like, enablement and giving people freedom to explore and start to build skills and everything, but also pushing them on [09:13] on we expect kind of throughput increase, we've seen a big, big increase in the throughput of pull request to our system. And, you know, like last year, like our CI system completely broke. It melted, you know, but I mean, it got like 10 times as expensive. And, you know, we did the work. We fixed the bottlenecks. We improved the performance of our CI system. That stopped being the bottleneck. And now CodeReview is our bottleneck. [09:43] as we did compared to nine months ago on our engineering team. And like, we're very proud of that. And, you know, now it's like, why can't it be 10x? [09:51] So what I love about this chart, just for a moment, is I had... [09:57] spent the last two decades of my career in product and engineering, last decade of my career as a CPTO. And it's so funny. I want to go back to a couple of things you said, which is one.

[10:07] You have to treat your org like a product. And I always thought that my job was not just the product strategy and the capital P product that we were delivering to customers. It was to design our organization to, I would say, like output innovation on demand, which is that was the job and less romantically, less romantically put. My job is to invest R&D. [10:30] for positive enterprise value. That was like fundamentally my job as a CPTO. And so what I love about this is it's merged PRs per R&D head. I'm presuming that includes, does that include product managers and non-engineering R&D or is that purely software engineers? [10:47] Yeah, this is all of R&D, and it's definitely the case that our designers and product managers and [10:53] TPMs, like every role in Intercom is really actively using clog code and shipping code and all that. [11:00] And also, we've been hiring. This number has not been static. So the number of PRs, the raw number, is dramatically higher than just 2x what it was a good while ago. So this is everything from your newest hire to your-- [11:15] product manager who's like adding some copy or shipping like small changes or whatever. You know, that's all based in this number. [11:24] The other thing I want to call out for folks is every board meeting I have been in for the last three years have said, [11:30] How are we getting well, actually, every board meeting I've ever been, period, has been how can we get more velocity out of R&D? [11:36] Certainly in the last three years, it's been how is AI inflecting

[11:41] our velocity and it's so funny I talk to so many people that are like it doesn't really inflect velocity we're not actually becoming that more efficient I'm [11:49] Is that true? Because I look at a chart like this and I say, this is a little bit more of what my instinct tells me is possible, which is if you go all in. [11:58] If you prepare your team, if you prepare your code base, if you have, as you said, I think a high trust [12:03] culture, people are going to look at this and say, oh, they're shipping these smaller PRs or engineers are gaming the system. I just... [12:10] I have not worked at a place that has such kind of like bad culture that that would actually come as an outcome of setting some sort of ambitious fun target like this. [12:20] And so I take this as at face value, and I think, [12:23] how is this not happening in your organization? Like literally the physical limits of my ability to type code. [12:31] are unlocked by by ai you should get some inflection there and so [12:35] You know, for VPs of engineering, CTO, even people that are on these R&D teams look at this and think, you know, this is possible and it may be a crude measurement, but it's possible. [12:46] I think, an appropriate one as a leading indicator of what's happening in your org around AI. [12:51] Yeah, and we support this with not just... [12:54] telling people to move faster. We're really looking from first principles of how to [13:01] how to do the work. We believe that all technical work will become agent first. And I'd like to set like a deadline for that, that at the end of the month, we're just going to go all in and it's never going to be the first thing that happens, say, in response to an alarm or in a planning meeting, that there isn't an agent in there kind of doing the basic work. And I think that's a realistic expectation, but it involves not just, you know,

[13:27] We're not just moving faster for the sake of it, we're seeing [13:30] that we're moving faster by looking at the fundamentals of where we're spending our time and reimagining how that work could be done in an agentic world. And honestly, if the agents didn't get better, if the models didn't get better, the harnesses didn't get better, we've got the building blocks just today to be able to just continue going, moving around, looking at how we do our technical work today. By technical work, I mean everything and delivery of product. [13:57] and move it to entirely be agents first and allow us to move [14:02] up to a higher level, to be able to work on higher level concerns, or just getting more stuff built, more stuff out there, or higher quality. That's all within every org's grasp today. But you have to [14:13] be very open to change. And I guess what's been fortunate to end to come over the last while is that we have been extremely open for change, both in [14:19] the product side of things and adapting the company to how I think companies need to work now with AI. And we're starting to see results. [14:28] Yeah, the other reflection I have upon looking at this chart is we're recording this in kind of the spring of 2026. Yeah. [14:36] Anthropic just said that they crossed $30 billion in revenue, I think up from $19 a couple months ago. And I suspect their revenue chart looks a little bit like your merged PRs per R&D chart. So how are you all thinking about... [14:53] the trade-off on cost here, right? Like, we're all consuming CLAW tokens. Yes, yes.

[14:57] you know, efficiency or output is going up, throughput's going up. [15:01] But is cost scaling proportionately? Are you all worried about is that the problem right now? Are you even worried about it? How do you think about that? [15:08] Yeah, we're definitely worried in that the build looks exactly like this. And, you know, I've spent a lot of my career worrying about AWS costs and worrying about our margins and stuff. And then suddenly you've got these costs showing up and they're disproportionate to the growth that we've seen anywhere before. It's like hiring whole new offices of people. And, you know, [15:34] But at the moment, our attitude has been, look, everyone just turn on Opus for everything. One millionth with a context window. You know, we just use the API plan, so it's all just on demand. And... [15:47] We think that there's enough... [15:50] alpha or benefits in [15:53] at this point, going as fast as possible and caring about the bill later because of the later benefits we'll get. Maybe that's a position of where Indicom is. I don't think it's realistic or feasible for absolutely every single business to do it. And honestly, I do kind of [16:05] respect, [16:06] when you have to actually think about your token use and how that can kind of force you to be more considerate. Or it sometimes even gets you better results. You know, you don't need Ocas for everything. Like there's faster models out there. And so [16:20] We're just kind of avoiding that optimization phase until the point of where we until [16:25] You know, we've gotten [16:27] serious benefits from investments in this platform. And so I think this investment, and I think it's, it, we are treating it as like an investment at this point, and is worthwhile. And both

[16:38] If this keeps going at this race, yeah, we should all work for Anthropic. [16:43] I think the way they're hiring, we're all going to end up working for Anthropic. Okay. And then one other thing, because I think folks are going to look at this, certainly engineers, and they're going, okay, you're shipping more PRs, but it's all slop. It's all garbage. Yeah. [16:56] You know, I know you all are measuring quality on the outside of this, on the other side of shipping all this stuff. So how have you seen this inflect? [17:04] your measurements around quality or customer value or what you're trying to achieve at the end, not just lines of code. Yeah, I have a stand alone graph that I can share, which is kind of interesting. [17:15] And so we've started to look at [17:18] the [17:19] the time it takes from the first line of code written in a feature to the time it gets posted on our news channel, like our updates. And that has decreased consistently over the last few months. And we're not optimizing for this, but we're interested in it. And the other thing is the sheer volume of things we have shipped also appears to have kind of just rapidly increased [17:49] We believe that these numbers, this increase in volume, is being borne out in real features, real products that our customers are using. And even we've been running some experiments, like how far can one person-- [18:02] get on their own building something that's plausibly a whole entire product area feature to be able to sell. So this is something we're taking seriously. And it's,

[18:15] We also care a lot about quality. We've been working with a research group in Stanford. We've been giving them our data and, you know, mostly just looking for any kind of insights to make sure we're not blind. You know, I join absolutely every single incident. I'm an ambulance chaser and I make like, and I'm not seeing any increase in kind of regular kind of incidents or editors or customer facing problems. We've had a few kind of weird problems, but not related to production. And, and [18:42] But also the interesting thing from the Stanford data when we checked back in on it last week was that their measures of, [18:50] code quality reckons that the code quality was improving. And, you know, the models are improving, the agents are improving, we're adding more and more guidance and skills and all these kind of things, which I think do craft or do force people down a road, which should result in higher quality output. But it's great to see when tools kind of can independently pull that out. Now, devils in details, you've got to go into the weeds, you've got to actually really have a strong [19:20] seeing some of the things that people are worried about out there. But that's it. We've got a mature environment. We're a 15-year-old SaaS company. We've been doing this for years. AI, I'm speeding up [19:30] your velocity will magnify all of your strengths and weaknesses. And thankfully, I think we've got a lot of strengths on the software delivery side of things that we've been able to take advantage of. One thing that I want to kind of call out here, which is you said that you've seen your code quality increase, which again,

[19:47] intuitively, I've always believed to be the ultimate end game of this. And every engineer, not every, many engineers that I've talked to just don't believe it to be true. But when you have the capacity to take on [20:00] tech debt, when you have the capacity to take on [20:03] the dragons in your code base, you actually can do those things, whether it's developer experience, security and compliance, just general maintainability of your code base, flaky test, improving your CICD. All those things become very tractable, not just technically, not just can an engineer execute on it, but actually the business. And I feel like people don't appreciate this. [20:26] The Business, capital T, capital B... [20:29] only has so much capacity for internal projects, meaning we can only allocate [20:34] so much of R&D towards [20:36] improving code quality. [20:38] Just how we live. We don't generate ARR on code quality, unfortunately. But when the... [20:46] When the cost of doing that compresses, [20:49] then you're able to say, yes, as a business, we should. [20:52] invest there one because we can and two because it'll unlock velocity on the outside and [20:58] for our agents and for our product managers and for engineers. And so I think this is actually a really important moment for folks to, [21:06] invest in code quality. And I often advise [21:10] a lot of CTOs and VPs of engineering when figuring out how to get their engineering team AI pilled [21:15] Say: [21:16] Everything you hate about the code base, go spend a month fixing and see how fast we can speed run that. That's going to feel really good. Okay, we've chit-chatted. We've shown graphs.

[21:25] point of how I is to actually ship some code. So let's switch over to that. We can probably come back to all these topics. I think they're so interesting. But you're going to show us how you all again in your mature code base, mature organization. [21:39] We're actually getting things live and some stuff you've done in the repo to make that possible. [21:43] Yeah, sure. [21:44] I'm going to do a fairly trivial change in our majestic Ruby on Males monolith. So this is millions of lines of code, all the tests. Yeah, it's the code base is older than Intercom. It was created before Intercom was incorporated. And, you know, it's got its problems, but we love it and we tend to it. And so I'm going to do a very trivial change in our majestic Ruby on Males monolith. [22:07] I'm just going to do a relatively simple change of adding a lobster emoji, Rails redirect to chatprd.ai. So also, I try and give hints to Claude when I'm actually demoing something. I don't know if it actually helps, but it makes me feel better. I'm just trying to add a bit of urgency here, you know? I think that's everybody's prompting strategy, which is, I don't know if it helps, but it makes me feel better. [22:36] Totally. So that's a nice way to interact with the agents, you know? [22:41] And so what we're seeing here is, I mean, it's already kind of figured out where to put a redirect. It's got the nice lobster emoji. And it's asking me if I want to open IPR. [22:53] So obviously I do. And [22:55] I think it's actually gotten the URL wrong. It's app.intercom.com, which will have the URL, but we can tell Cloud Code later on about that. So what we're seeing here is

[23:06] First of all, an important point. I'm just going to scroll back up. [23:09] One of the things we noticed early on when we started getting Cloud Code to write all of our code, and we're up well above 90% now, is that it would create pull request descriptions that were kind of terrible. It would describe the code, and that's the least interesting part of a pull request. You actually, as a human, or even as an agent reviewing code, you want to know the intent behind the pull request. You want to know the interesting bits, what's kind of related to this, and... [23:38] You know, LLMs are very good at just regurgitating or rewriting code into English. That's fine, but it's not what we need. And so one of the things we noticed as well, when people were using Cloud Code, we created an LLM judge to evaluate, because we had suspicions that the quality of the pull request descriptions was going downhill. So we created an LLM judge to evaluate [24:00] What does a good pull request... Well, we decided what a good pull request description should look like, and then got a Nellander to go through all, like, months and months of data. And yeah, the trend was awful. The trend was going in one direction. And this is bad. And, you know, look, humans aren't perfect at creating pull request descriptions. Sometimes they're just blank and whatever. But... [24:25] I think with our use of tools like Cloud Code and setting up these kind of platforms around us, you really have to be pushing for higher standards. [24:33] as close to perfection as possible. And this was clearly something that we're just not going to tolerate a lowering of standards in our environment. So we created a skill called CreatePR. And what it does is it uses whatever context it can from the session to

[24:48] describe the pull request. [24:50] It's not quite rocket science, but often the session knows exactly why it's doing the thing. And so, but then we had to kind of force it in, you know, we started, you know, we started, [25:00] We told people like, oh, just use the create PR skill. And then people would want to use it. You don't really actually want to have people remembering things. So we added it as a hook. So if Claude decides to, you know, use the GitHub CLI to open a pull request, we just block it. And we say, yeah, tough. You need to use the create PR skill. And also, you're probably going to have to figure out a different text description. And then I might interview you if it's not enough context there. [25:30] - Yeah. [25:31] There's enough context in this. But the point being that this is a platform, we want great outcomes, and we measure the inputs and outputs. And after we put this in place, the LRM judge [25:44] Reckon we're doing a great job now. And so we're at higher quality pull request descriptions now. [25:49] Now, this is not the most important thing in the world. This is not going to get intercom to 2x or to 10x revenue or anything like that. [25:58] It's all of the composite little jobs that when you assemble means you have an extremely competent engineer who works appropriately in our environment. And that's where we're putting our investments for each little skill and hook to do these things. So they almost look inconsequential, but they result in better outcomes. And so we look through here. It's creating a PR. I'm going to have to check on what it's going. This probably will be automatically approved as well, which is pretty cool.

[26:28] be back as well in action. Ah, still building. We'll come back to it in a couple of minutes. [26:33] One thing I want to call out for folks as you were describing sort of why you put in this skill to improve the PR. And for those who don't know, a skill is basically just like a set of instructions and sometimes scripts that... [26:47] a LLM or a agent harness can invoke at a certain step in your flow. [26:52] One of the things that I was thinking as you were describing [26:55] why you put this skill together and got really opinionated about PR descriptions is in engineering, we have been able to architect really opinionated CICD pipelines. [27:06] So how written code goes from being written to deployed in production. And we have, I mean, you saw it in GitHub. We have all these checks and lints and pre-deploy. [27:17] you know, preflight things and preview branches, all these things. [27:21] Once the code is written, [27:22] What I think is really interesting about skills is you can bring some of that determinism to... [27:28] as you write the code. [27:30] how you want that process to go. And we used to not be able to do it because it used to flow through the hearts and minds and hands. [27:36] of humans, which are much harder to put in these structured guardrails. And we would do this by writing wikis or having, you know, SOPs where it said, can you please follow step A, B, C, D, E, [27:48] And now you can just make it really easy to enforce those standards across the team. [27:53] Which I don't think is... [27:54] micromanaging, it's actually just making everybody's golden path much smoother.

[28:00] to production. And so I think there's this just very interesting parallel to how we've approached CICD [28:06] to how we approach things more upstream, even from the product management perspective. [28:10] Totally. [28:11] We're on this movements towards a software factory. [28:15] And [28:16] What factories are great at is, you know, like an Ikea factory or something. It's all the same furniture, all the different bits, and you know how to assemble it. And look, it's not your artisan stuff. It's not, or it's not cutting edge or whatever, but it's very predictable and, you know, has a certain quality and meets certain standards when it comes out the other side of the factory. And so, well, pull request descriptions. [28:39] Again, they're not make or break for the factory or the pull request or whatever. It's one of those qualities of just good quality work that's reliable, predictable, and then when assembled together, [28:49] You've got your IKEA factory. Well, and people don't want to feel, certainly engineers don't want to feel like they're part of a slot factory, right? And so these things that you can add into the flow that actually up level and meet the standards of the engineering team, [29:03] really help your human engineers on the team feel like they're working in a place that values quality. And so I appreciate that you've put that effort into [29:14] into these behind the scenes hooks and skills, [29:17] Because I'm sure it reinforces to a culture that's being asked to move very fast, to ship things differently than they have before, that you still do care about their experience. [29:28] reading pull request descriptions, you meet their bar for quality. And I just think it makes everybody happier.

[29:38] Yeah, well, it's great when the robots just produce the work that you'd expect of your best engineers. You know? Yeah. And, you know, maybe as you get this live, I also think there are just still such [29:50] more interesting problems to solve in software engineering. And we can talk a little bit later in the episode about some of the interesting problems that you are solving on the product side, on the technical side. I think there is no lack of hard, intellectually stimulating, creative problems to solve for customers. And coding redirects is just 100 percent. [30:10] Not one of them. [30:11] So did we get my redirect live or are we close? It's still there. I'm waiting for an automatic review to kick in, but we can come back to it. So one of the things I would like to show next might be some of the telemetry that we have in place. So we saw that, you know, there was different skills getting invoked and we're going to be [30:33] We don't like flying blind. [30:36] To run a system like this, you need to know how well people are using it. Are people using these skills at all? You know, the kind of basic information that you'd expect of when you ship a product to your customers. Like, you know, where can I see the usage? How can I fight for the usage? What's going wrong or what's not going wrong? And so we collect a bunch of... [30:56] telemetry using different mechanisms, and have different homes for us. The most open one that we have is we collect basic usage information for skills and the like, and we send it to Honeycom. So we just have a shared key that's deployed to all of our laptops. And anyone can go in and kind of look through this data. So if you're developing a skill internally in Intercom, and like hundreds of people do this, it's very easy for you to go into discover like, hey, who's actually using this? When are they using it?

[31:26] And you've got to use this as a kickoff to follow up on just like basic discovery of usage of your skills and all. And unsurprisingly, the main skills that we have are things like creating PRs. Admin tools is our admin. [31:40] like internal tooling APIs, where we have an MTP in front of it. BuildKai is our CI system. Snowflake Logs is where we put Snowflake. So you can see from this, a lot of work, a lot of the skills are being involved are all around the building, and then seeing where my stuff is, and maybe some troubleshooting type information as well. And so this is the first step. It's like, if you don't have this, it's hard to have a large system. [32:02] like [32:03] all these hundreds of skills and hundreds of creators working in this area without having decent telemetry. The next thing we do as well is we also collect all of the session data and put it into S3. And so we anonymize it. We do a few things to make sure we're not doing anything too private. You know, people put all sorts of stuff in their sessions. They yell at their sessions. Yeah. Yeah. People have personal relationships at times with Claude. And like, we don't really [32:33] into how things are going. You know, I think understanding [32:38] Like what the dropout rate of sessions like, dude, [32:41] how quickly people got to something useful, like whether it was a PR or something like that. This kind of information is pretty interesting. And so we're harvesting a lot of session data, we're doing different things. This is what I'm showing here on the on the screen is like a very simple tool that we put together, which just gives you some personalized insights. And, you know, you can do this inside cloud these days as well. There's there's plenty of skills out there on GitHub where you can do session analysis. But

[33:09] I think we just built a little tool on top of our session collection system to give people feedback. And it's feedback that we're interested in giving feedback about how their sessions are going and how they're kind of fitting in, how you should think about your own, I guess, use of CloudCo compared to everybody else in the org. And, you know, I'm not doing too bad here. It's like 79th percentile. You know, someone has to be down the bottom of every percentile. And [33:34] There's some interesting feedback here. [33:38] It's kind of getting annoyed at me. Oh, Ron, I was getting annoyed at Claude a few weeks ago because I'd set up GOG to interact with all of our Google stuff internally. And... [33:50] But I kept on trying to do the wrong thing. And I was kind of giving out to it and ended up adding stuff to Claude and MD and stuff. [33:57] It's kind of giving out to me here. It's reminding me that this wasn't a very effective way to interact with Claude Kerr. So, you know, it's a good prompt for me to actually go and fix up my memory or whatever. And like we all like people are at different levels, even at Intercom. People are at different levels of adoption. People are joining Intercom. They may not have like seen a system like this before. [34:20] So, [34:20] how things are going and get feedback. So this is one example of how we're just trying to pull together this information to give useful, actionable insights to people, so that they feel supported and that we're not just throwing them an API key and saying, "Best of luck." It's like, no, we understand what growth looks like and the progression that people go through when they're using these tooling, and getting better and kind of self-improving, and we want to support all that. So this is one of the things that we're doing with the session data.

[34:50] like being able to, like we want to [34:53] get insights to which skills are the highest quality, which skills get you to your results as quickly as possible, and then which ones need work, which ones aren't working out so well and might need a bit of attention to improve. [35:08] This episode is brought to you by Cursor. If you all have been watching How I AI, you already know this. Cursor is my favorite way to code with AI. Whether I'm using plan mode to build out an ambitious feature, reviewing AI generated diffs right in my editor, or kicking off cloud agents to multi-thread our roadmap, I reach for Cursor as my favorite multi-model coding platform. Even better than building myself in Cursor, I love collaborating with BugBot to fix PRs for [35:38] and have begun relying on Cursor's automated agents to keep our codebase clean. It's not just me. The most ambitious teams love Cursor 2, including engineers at Stripe, [35:49] OpenAI and Figma. Ready to build more? We're giving $50 in cursor credit to HowEye AI listeners. [35:57] Claim your credits at chatprd.ai slash howiai. That's $50 in cursor credits by going to chatprd.ai slash howiai. [36:09] I have to pause before we look at your list of skills because I'm so excited about that part. But if folks aren't watching it, they may have missed how amazing what you just showed is. So I'm going to reiterate it, which is.

[36:20] One, you've instrumented. [36:23] all your internal skills, [36:25] With telemetry, so that – and you're using Honeycomb. Love the Honeycomb team. You're using Honeycomb to see – [36:32] how often those skills are invoked over time. [36:35] So this is just a tip for anybody building out a skills repository internally or even somebody who is maybe trying to get. [36:44] some visibility into their impact across the org. Let's say you build [36:48] a skill and you want to go to your boss and be like, boss? [36:51] My skill is being used by literally everybody every day. [36:55] Find a way to put event level telemetry [36:59] invoked in the skill a little dashboard and you can track those over time again treating your org like a product [37:06] treating your repo like a product treating your ai setup as a team like a product and all products all good products have tracking plans and so figuring out how you put that telemetry in i think is really smart [37:18] And then the second thing for those that missed it or how to do it is you're taking all the raw session, I'm presuming JSON files. So for folks that don't know, Cloud Code stores all your chats with Cloud Code. Yeah. [37:31] on your computer in JSON, and you can go look at those or query those at any time. It sounds like you all are uploading those files to S3, and then layering on top of it some anonymization, some user-level views, [37:45] And then you're essentially building sort of what I would call like an internal eval of

[37:51] how people are using Claude code and what problems they are having over time so that individuals, one, can triage their own implementation. As you said, oh, it looks like I need to do this or that or improve my agent's MD. Right. [38:05] But then if you're seeing consistent themes over the organization on, you know, [38:11] it's never invoking this mcp when we need it to invoke this mcp or people are yelling no every time the create pr um skill gets queued up you can fix that at a systems level but you can't do that if you don't have the visibility so again [38:25] my VPs of engineering, my CTOs, my friends out there, [38:29] Put some telemetry in your skills. [38:31] and then do some meta analysis on your Claude Code sessions across the org. [38:36] And you'll be able to identify places where some probably some high leverage fixes are going to get your team unblocked over time. [38:42] I do hope and expect that this stuff will get easier over time, you know. [38:47] I'm happy to kind of invest the work so that we can move fast and kind of be on the bleeding edge. But there's something to be said also for having like last mover advantage and just getting all this stuff for free whenever Entropic ship is or whoever ships. I mean, maybe this is a product just that people should buy or build. [39:07] um but for us right now we've no choice we just gotta build it we're we we're we're we've [39:12] We're fascinated with the insights that are locked away in these sessions. And so we just got to build stuff so that we can see what's going on. [39:20] I love it. OK, can we see some of these skills?

[39:24] Yes, so... [39:27] It's a very exciting GitHub repo. Our lives are all GitHub repos and Markdown files. Totally. And [39:36] We have a lot of activity at the moment. We ran an AI day last week, kind of getting more people, contributing to us. And so, well, so what this is, is it's a plug-in [39:48] uh, [39:49] repo and we have a series of plugins and they've been they're growing daily at the moment and kind of every team will have their own kind of specific plugins and actually in general though we're very liberal we want stuff to end up in here even if it's not great and but we do sweat the details on the core plugins things that we think are fundamentals foundational ones that go out to everybody and so [40:13] Where we start off was we have these base plug-in, which gets installed. Oh, yeah, so we distribute this not via the code plug-in mechanism. [40:24] It was just a bit flaky. It was... [40:26] Sometimes it updates, sometimes it wouldn't. And it ended up kind of like trying to manage a Python install on hundreds of different laptops. You just don't want to do it. And so we ended up using our internal IT systems to synchronize all of the plugins [40:41] to the disks of everyone's laptops. So this is a great cheat code. And yeah, strongly I recommend getting very close with your IT team to be able to deliver things like this reliably and not have to rely entirely on the cloud code plugins mechanism. Just our experience is a bit flaky, and it just gives us a lot of reassurance. We don't have to do certain types of debugging once it's all on disk. So we know this stuff works anywhere because we've got our IT team pushing it out to disk.

[41:11] hooks. We have some of the foundational things like merging PRs. We don't want our agents going off into AWS. And then just different settings and the telemetry things as well. So these are [41:23] the core things that absolutely everybody gets. But these are minimalists. We don't want anything that could be inappropriate in, say, a non-technical person's laptop or whatever. So this is the basic building block. The next main bit for us is what we call developer tools. [41:44] Again, this would be... [41:47] things that we then do all of engineering and beyond at this point. And these would be generally skills that would be appropriate to be used by any engineer in the course of their work day to day. And again, we would have a high quality bar again for all of these. These would all require evals. These would all require to pass different kind of tests or analysis that we do on the quality of skills. And so we try and maintain these and make sure that they're well updated and well used and we pay a lot of attention to. I can [42:16] maybe go through one of these skills in a bit of detail. This one's near and dear to my heart. It's flaky specs and [42:24] I think the interesting part here is not... [42:27] the skill itself. The skill [42:29] does reliably fix flaky specs. I can pull up in the meantime, [42:34] Like, here is a list of flaky specs that we have at the moment. [42:37] I'm going to open up the skill and just start to run it on this issue.

[42:43] And so while this is running, just walk through what's in the Flaky spec. [42:47] skill. And so there's a checklist here. And the fun part about how I built this was not that I'd be [42:54] was a world-class expert at fixing shaky specs. I roughly know the problem and have fixed a few of them in my time. But there's different class-- in a large test environment like ours, we have hundreds of thousands of tests. And if you're not super careful about data poisoning or race conditions and all these kind of things that kind of kick in when you're running, [43:15] millions and millions of tests a day. You end up with these tests that end up slowing down. [43:19] your ability to deliver code to production fast and reliably and not confuse developers by things randomly breaking. And there's kind of known patterns and known ways you would go about this. [43:31] My goal. [43:32] which was to have... [43:33] a skill fixing all of these flaky specs. And it was something that agents are pretty good at when you give them a kind of testable goal. You know, this wasn't quite open-ended. And I also had this huge backlog or yeah, there was a backlog of probably a few hundreds, but then also all of this historical flaky spec information. And so you can just [43:51] harvest all of this data in your environment to go, hey, Claude, I'm going to build a scale. [43:57] first of all, go and research every single flaky spec we've ever had. And then we're going to build a checklist. We're going to build a mechanism. And then we're just going to crunch through them [44:05] over and over and over. And you get to this 1x kind of-- it's doing a good job, probably as good a job as I would do. But then as you keep building up all of these little teeny steps, which are the kind of things that our best Rails coders kind of do, they've got all the stuff in their heads, and all the different classifications of flaky specs, and verifying it gets real data and--

[44:29] But the really fun part is then, you get something that's starting to be like 10x. It's fixing flaky specs that [44:35] I'm not even sure if I could do. It might take me a day or something. And I probably wouldn't do it. But then you start to add in stuff into the [44:43] the skill along the lines of like, OK, when you fix something and it's novel, you need to update yourself as well. So in that session, it's updating the skill. So the skill itself is kind of learning as it goes along. And we also fan out. So it's like, OK, I'm very happy that you fixed that flaky spec. Now, find every flaky spec that got impacted by that nature of it. [45:05] I went from zero to like 100x in terms of [45:09] This skill now is like, [45:11] you know, see your distinguished engineer, or being able to fix these specs. But it was more like the process that got there. And so like working with a feedback loop, [45:24] working with a very clear goal, and then giving it the freedom to do it, giving it access to the systems where it needed to pull in metadata, be able to run builds itself, and having that feedback loop where it's learning. And then designing the skill as well so that it's [45:38] have to edit it every so often. It ends up taking up too much information that might confuse things. But then you break things out into like reference guides. So you're doing this like progressive discovery thing. And I've even accidentally pointed this skill at like a Python code base. And Claude has just gone, "Eh, it's just Python. I'll be able to go." And it kind of uses the knowledge that's applicable to it. And so again,

[46:02] This skill is not going to make Intercom's revenue go 100x, [46:08] It's now this, like... [46:10] perfectly reliable, [46:11] thing that we really no longer have to think about. Now, we can expand out into many, many different areas. [46:17] We just have to maintain this. And the maintenance work for a skill like this just isn't much. And we have... [46:22] evals and stuff so that when we're upgrading models, or maybe even moving to cheaper models or whatever, that we can make sure, yeah, this thing isn't progressing. It's still working as well as we think it is. And we've got confidence and certainty that this is still a reliable building block. And again, the constituent parts put when put together, you've got like a very senior engineer who's able to get any work done in your environment. And so yeah, we can take a look at what it's doing. Oh, it's asking me for permissions. [46:48] You forgot to make no mistakes, dangerously skip permissions. That's the rule on how AI. One thing while running, I wanted to say is, you know, this skill is a perfect example of what I call the like, and then. [47:02] AI workflow, which is [47:04] i tell everybody like pull your skills and pull your workflows through a bunch of admins so [47:11] I want to fix flaky tests. [47:13] So I go to GitHub, I find a flaky test, I run through the skit. Let's say you fix it. And then what would you do? [47:20] Well, I would document how I fixed it. And then what would you do? Well, I would go find all the other ones that are just like this and fix that. [47:27] And then what would you do? [47:28] I would go from [47:30] you know, our Rails code base for a Python code base and apply the same. You could just do that over and over. And because the cost of running these is so low, you can actually pull the thread of a bunch of stuff.

[47:42] any reasonable human would have quit at step one [47:45] Because you're not limited, again, by headcount or coordination costs. You're limited by the technical capacity to solve the problem. [47:52] which I think is a really interesting way to think about [47:56] how you get from like, [47:57] The, you know, engineering intern that whose job is to go through and take a first, you know, gentle pass at all these flaky tests. [48:06] through to the distinguished engineer who has just speedrun through... [48:10] 300 of them. [48:12] and has thought of a completely different way to architect your testing overall in your repo. [48:17] So I think that's a really great [48:19] model for things and [48:22] And then the other thing is like, again, [48:24] Engineers, go speed run your tech debt, fix your flaky tech. Like these are all things. [48:29] that as somebody who has run engineering organizations, [48:32] I have heard over and over [48:34] We can't because our code base, blah, blah, blah, blah, blah. Like, can we pretty please allocate this amount of time to just fixing this really annoying front end flaky task? [48:44] You don't have to ask permission for that stuff anymore because there's just a new way to solve it. And I think, again, just going back to some of the stuff we were talking about earlier... [48:54] I think your overall product quality is going to go up. I think your overall developer experience is going up. There's just so many good things that come out of using these tools and using them correctly. [49:04] Yeah, I think backlog zero is a realistic thing for teams to be able to go after. You know, all the things that you wish you'd [49:11] ever wanted to do.

[49:13] you know, it's now just achievable. But of course, you've got to balance it with, you know, all of the extra stuff that you can just deliver at the same time. But it's so sweet to be able to think that, hey, we actually have a path to getting rid of [49:24] all of our backlogs and all of the kind of architecture changes or whatever, you know, we can [49:29] Recently, I was... [49:31] taking a Go microservice and re-implementing it in Ruby. It was a single cloud code session. Before November, this was something that I would have had to advocate for on a roadmap and plant some seeds in different engineers' heads and build people towards it and blame a lot of problems on the existence of this microchip. Wait, trigger warning first before you talk about that process. [50:01] Influenced.org. But yeah, but now it's like, well, I don't even have to think about this now. It's a single session. And in fact, I can... [50:08] I can get Claude to implement it five times and compare the styles or compare the, you know, get it to review them and figure out what the best way of implementing the thing is. And this is just like this level of kind of creativity and freedom that where like your imagination is the blocker, not, not. [50:26] the time it takes to actually knock out one of these things, which was months in the past, you know? I completely agree. And I feel this at Chat Parity where people are like, what are your, I mean, I'm a product tool for product people. They're always asking what my roadmap is. And I was like, I literally don't have a roadmap. We burn down the roadmap every week and then we figure out what we're going to ship next. And of course we have thematic ideas we want to pursue and things that are larger. And one of the things that I do

[50:52] to keep myself from [50:55] over shipping absent product market fit. [50:58] is literally constrain the ideas to what I can do in my brain, which is there's like a natural throttle on not getting slop out. [51:06] Because it's not engineering broadling me. It's actually just good commercializable stuff. [51:11] ideas. [51:13] And I think that's where we're going to see some of the limits start to come in play. Again, referring to Anthropik. [51:19] Another big news piece came out is that they're hiring a bunch of PMs because they have so much engineering capacity. [51:24] They're actually limited at the PM capacity. And so it'll be interesting to see where the bottlenecks go. [51:30] in your business. [51:31] you know end up and which bottlenecks are appropriate it's probably good to have a product bottleneck a little bit [51:37] because then you're not shipping [51:39] anything [51:41] which customers can't absorb. And so I just, I think it's going to, and it's going to evolve over time. And then, you know, product is going to have a whole set of skills and then [51:50] I don't know what we're going to do with our time, hang out on the beach. But I think it's a pretty interesting time to run orgs. [51:57] Yeah, you know, I think engineers, designers, product managers, maybe it's just all going to be one blob of builders or something like that. Let's do it. Everyone, everyone just does things. Everyone just does things. Yeah. And, you know, it's great. It's lowering the barriers to like just getting a lot of stuff done. And it's like so much fun when you can, when you don't have to ask somebody or get something on a backlog or whatever.

[52:27] very fast in a small group. It doesn't matter what your discipline is. It's just like a great leveler at the moment. [52:32] So yeah, so we're live. I think our logister is live. And it should be on app.intercom. [52:37] lobster emoji. Look at that. It was amazing. I need to get you all an affiliate code, you know. [52:44] Uh, yeah, I mean lobster emojis, they're the new thing. They're the new growth hack. They are the new growth hack. [52:51] Okay, so we have seen... [52:53] Your PR per R&D... [52:56] employee [52:57] Go up. [52:58] We've seen how you can get from kind of Claude code to production very, very fast with a bunch of guardrails. [53:04] We've seen your list of it looks like hundreds of skills, but at least dozens of skills that you're invoking. [53:10] via hooks, [53:11] You're using that to not only ship customer facing product, but you're also using that just to make developer experience better, burn down tech debt. [53:18] All those things we want to see. You all are you're measuring it both from a telemetry perspective. [53:23] both like quantitative and qualitatively. You're measuring your cloud code sessions. And, you know, 2x isn't enough. You're going to get to 10x. So... [53:33] You all... [53:34] are on the edge, at least for folks that I talk to. And I'm sure you're like me where you're like, [53:39] Sure, you think we're on the edge, but then I see people and they're really on the edge. So we always have ambitions to move forward. [53:47] My question now to you is, [53:49] How has this impacted how you think about your customer's product? You know, I'm an Intercom customer. I'm a thin customer. [53:56] I interact with Intercom Code and Intercom UI

[54:00] literally every day, my open claw. [54:02] has an intercom API key. How do you think about [54:06] Now that you have this experience with Cloud Code internally, how do you think about what that customer experience is going to look like? [54:12] Yeah, I think... [54:13] There's a few things going on. One is [54:15] that people are outsourcing a lot of decisions to their agents. [54:19] And this is a good thing in many cases, but there was good research done recently about, what does Cloud Code pick? And certainly I've had the experience in the distant past where I'd ask [54:30] an agent to [54:32] add something, except do it behind a feature flag. And then it would start to go and implement its own feature flag system. No, no, no. [54:41] in our code base, which has a pretty sophisticated, old school, home-rolled feature flag system. So, you know, nowadays, mostly we'll stick to whatever's in the code base and that's fine. But, you know, SaaS products, they're really good at their jobs. They're actually worth paying money for. And getting back to the feature flag, [55:02] situation, you know, if you're building a new business, you're relying on your agent to make decisions, often an agent will [55:12] When prompted, it's like, hey, how should I solve the feature flag problem? I want to make sure I'm doing all these safe deploys and that. The agent will just go, yeah, I'll do it myself. And the kind of [55:21] build over by decision. And you can see why the agents do it this way, because they can achieve this, they can get it done. They don't have to rely on the human, okay, like open claw changes things here a little bit and maybe computer use does as well. But still, we're not really, we haven't really adopted SaaS businesses to be agent friendly. That means,

[55:42] well, all sorts of things around how do we position our websites and content, and how do you get updated in their knowledge, and how do they discover it? But also, can they actually just get it done? Like, can you ask an agent, "Hey, could you [55:57] just sign me up to Intercom and get me, uh, Fian working on my website. And so, uh, like this goes along sides, just. [56:06] having to make more APIs for things. I think I'm kind of like... [56:11] omni-channel as such. I think there's a feature for CLIs and MCP and REST APIs. I think [56:18] I'd like us to get more comfortable around things like ephemeral APIs or multi-step APIs. I think CLIs are good at wrapping these kind of things. But the whole point of all this, where I'm getting at, is you want to be able to just help agents out at the time when they're interacting, they're in discovery mode. And you want to give them clues. You want to give them hints. You want to give them help to be able to do things like sign up for something fully without having to go back to the user and say, yeah, sorry, can't help you there. You've got to go away and figure out how to sign up for something. [56:48] I've been working on something in the last few weeks, which hopefully should solve that problem. And I can paste in a prompt and then see how far it gets. I also just, while we're running this, I have to go back to your feature flag example because it [57:01] You know where I used to work. It broke my heart. [57:04] That Build It Yourself was at the top of the feature flagging list. But I do think I have a paranoia.

[57:13] moment about this, which is... [57:16] Model providers and harness providers are highly incentivized. [57:20] to [57:21] Build it yourself. [57:22] consumes lots of tokens versus buy it. [57:27] Maybe consumes less. So I'm just really interesting to see how this all... [57:32] shakes out. You know, people people are very anti sass is dead. [57:37] And I'm a little bit more like, yeah, but like the current form factor of SaaS really is... [57:43] has something coming for it in a particular... [57:46] DevTools because these models are so good at writing code. [57:51] I think you're in a real [57:53] pickle to try to figure out how to find the right value wedge at the right moment, how you can allow agents to not just sign up and set up things, [58:03] but purchase it you know like what does your trial experience look like if your first user is an agent i think all of that is super important and then you know to your point earlier where you [58:16] APIs, ephemeral APIs, CLIs, MCP, I think the answer is yes right now, which you cannot predict. [58:23] the the medium by which a user is going to come to your site. [58:28] They could come through a search and hit your website and download things and look through docs. They could come through Claude code. They could come through an open clot. You just really don't know. [58:38] And so you sort of have to meet your customers and your non-human customers.

[58:43] where they're at [58:45] And I think it's really smart for teams that have any part of their product that needs to be implemented via code. [58:52] to be thinking about this problem yesterday because you will be left behind, I think, if your agent experience isn't there. [58:58] Yeah, I agree entirely. And I think there's a whole craft in how to make a CLI agent friendly. I think MCPs obviously get that right a lot of the time. But, you know, for example, one of the things that we do in the help is like kind of just give a hint to the agent. It's almost like prompt injection to a certain extent, except it's not malicious. You're just trying to get it along to what it's trying to achieve. It's like, well, maybe you could check email. And if an agent has access to your email. [59:28] I was looking at it. So it's just they're going, oh, yeah, I can probably get this done. Or like you can hint to them like, [59:34] I've kind of cheated with this. So this is my own personal website hosted in Vercel. And it is, I've kind of pre-populated a few articles so they can upload and Finn has some content to answer questions with. But you can also just return in the help going like, hey, you know, you should probably think about creating some articles if you want Finn to actually start answering questions. And that can be extracted from the code base or whatever. Yeah, being like-- [1:00:02] I also think a lot of interfaces, like CLI interfaces, like I use GOG. [1:00:07] It's part of the OpenClaw universe. And I think it's a lot better than the official Google GWS one. But I think if you start to use it, it's actually just more human. As in, the interface just kind of makes more sense to a human. I think the Google one is like, I kind of get what they're getting at. And there's kind of JSON in there and stuff like that.

[1:00:35] more human friendly or something things that are effective for agents can often be things that are more human friendly because they're discoverable, these verbs and words and not just kind of inscrutable weird stuff going on. And in command line options, I think I've confused Claude here. I'm not sure what where is that's OK. I'm going to I'm going to narrate for folks what's happening here, which is you basically said, like, install intercom on this site. [1:00:59] There's an intercom CLI that's like, cool, I can access the intercom APIs and do a lot of this. My favorite part of it, though, is signing up, getting a verification email in your email address, invoking via like this hint, basically, of like, if the user has email access set up in however you're accessing it, go check for this verification email because we have a code in there that we got to snag. And because you're using GOG, right? [1:01:27] which is a command line tool to access Google Workspace. [1:01:31] You can go do that. [1:01:33] pull that code in. And what I think is so interesting about that particular flow is that [1:01:38] You know, I think AI is creating sort of race conditions in shipping across the org. [1:01:45] which is like you can YOLO a CLI probably faster than whatever team [1:01:50] that manages email authentication can change how email verification works. And so you're like, I'm not going to let that happen. [1:01:58] break my product. [1:01:59] What I'm going to do is create a flow that I can use that sort of sticky part of the flow. [1:02:04] AI brains and get through it. And so again, your product doesn't have to be perfect for an agent to traverse it. And this is one of the things I'm actually really excited about SaaS is all those things that are just so complicated to do.

[1:02:20] as a human, multi-step forms and like, [1:02:24] nested fields on nested fields and finding, you know, categories and [1:02:28] Just those things that I would say UX designers and product managers have written their most tedious PRDs on. [1:02:35] and done their most detailed specs on, like, you don't actually have to worry about making that quote unquote usable because you can just brute force intelligence against it and and solve the problem. And so. [1:02:48] I think that's interesting because the core value proposition can get bigger and bigger without being constrained by the surface area of a website or a UI or any of those things. [1:02:58] And so I think if you're not thinking about what does that CLI look like for you, [1:03:03] And what adjacent systems... [1:03:06] Does your product butt up against? It may be email, it may be some other dependency. [1:03:12] And how an agent might traverse those systems and, [1:03:15] You're just going to get less and less adoption because this is going to be more how people install products. [1:03:19] Yeah. [1:03:21] If I don't poke holes and if I don't make a CLI that kind of bypasses some of the ways the product works, somebody else will. You know, they'll just put their own agents on it and they'll burn more tokens. They might get frustrated. You may as well shortcut them and give them an interface which just works. May not be the perfect interface, but that's the beauty of these things. You can get updated over time. Agents can just pull down the latest version. [1:03:46] And yeah, hopefully I have something to show here, though.

[1:03:51] I want to call out while you're talking about that, which is as I'm watching this and it's taking some time to build, [1:03:57] Your conversion rate drop-off point is somebody pressing the escape button. [1:04:02] And just saying, forget it. Like, this is clearly not working. What if we built it ourselves? And so I think it's a really interesting moment for product managers who are [1:04:12] right now are not getting the visibility. [1:04:15] of the drop off, right? When you were going through a website, you could put telemetry in it. [1:04:20] You could say, OK, users going to the sign up page, drop off, email verification, drop off, going to the docs, drop off. You could build this nice little funnel that identifies where your users are having problems. [1:04:31] You can put some telemetry in your CLI, but at the end of the day, some of that drop-off and the alternatives is very invisible to you here. And... [1:04:39] the switching cost quote unquote is like pressing escape and saying do it a different way [1:04:44] And so, again, how quickly you can speed run to a zero to one installation in an agent, I think, is something that. [1:04:52] Everybody should be running right now. [1:04:55] And it doesn't just have to be a code product. Like I think more and more [1:04:59] People are doing non-technical tasks and interacting with non-technical SaaS. [1:05:05] in [1:05:06] Claude Code, in Claude Cowork, [1:05:09] And so even if you're not DevTools, if you're not thinking about how can a user do this quickly, [1:05:16] in a third-party harness or system or an agent can do this quickly.

[1:05:21] um, [1:05:22] you're really missing out on customer growth. [1:05:25] Totally. [1:05:26] Okay. [1:05:27] How are we doing? [1:05:28] It's on its fourth attempt. That's fine. And you know what? Let's press the escape because you know what? Let me tell you how cheap that exercise was. It was like five minutes and some tokens. And you're going to spin up a fresh Claude code. [1:05:45] I don't know if you put make no mistakes. That was probably... [1:05:49] what we missed. Make no mistakes. And it could have done it. And again, this is just learning. Like, why isn't every engineer, every PM doing this? [1:05:59] once a week or once a month just to figure out [1:06:02] how it can work. I think it's great. So [1:06:05] Right, and you've shown us everything. You've given us all... [1:06:08] All the secrets. Let's [1:06:10] Get out of the terminal. [1:06:12] And let's do some lightning round questions. [1:06:15] So my first question for you is... [1:06:20] How does it feel? Because what what I observe from our conversation is it feels fun. Like culture has, in fact, gotten better, not worse because of this investment. And so, you know, as a company that has really put in the effort, both on the on the customer side, [1:06:37] and internally. [1:06:39] How do you think it's shifted culture? Has it at all? What have you observed? [1:06:43] Yeah, everything is just faster and more exciting. You know, I mentioned feedback loops a good few times, and [1:06:50] you know, you can just get stuff out there so fast now. And I've been having the most amount of fun in my career over the last three months or something like that. And like, it's fun in many ways. It's fun because I can do stuff that, again, I would have had to convince other people to do, or they were just things on my wish list and I could never get around to them. I just kind of complain about them. But now they're just realizable. But also the fun aspect of like,

[1:07:15] making other people productive, like leveling people up, like removing work. I had like, [1:07:21] Intercom is a pretty good culture around resisting [1:07:26] like the kind of slow movements towards being a large company and all this process and stuff like that and we're kind of in denial that we're like a large company um i think it's a healthy way to work in many ways and uh but this has kind of got us back to our roots in in a loss that's you know you can make fast decisions and get them delivered and get that feedback super fast um and i've been able to like ship actual features like not just the cli but i ship ship some webhook features and [1:07:56] It's been a long time since I've done that. I've been in the weeds in platform space for a long time. But it wasn't even a big deal. It was just a couple of hours, just kind of get something done. It was something a customer asked for. So my job has become more varied. [1:08:09] I'm able to kind of see more and get more done and help other people get a lot more done. So you get this kind of excitement and velocity increases. And, you know, we have all those measurements and that's all kind of good stuff. But just the excitement of waking up in the morning going like, I'm going to get a lot done today. Like that is a fun way to go about your day. I completely agree. And I hear this over and over and over again. I certainly feel it myself, which is this is the it brings me back. [1:08:35] to why I learned to code. It's like that same moment of, [1:08:39] I didn't learn to code because I like to type code. I learned to code because of the magic of you running code. [1:08:43] like hello world and it it shows up somewhere and that feels so it's just a very creative experience which leads us to my second question which is

[1:08:52] I see all the time that... [1:08:55] One of the most impactful change agents inside an engineering organization can be a senior principal engineer saying, let's go. [1:09:02] ham on some AI code. [1:09:04] And? [1:09:05] The single most blocking person in the organization can be a senior principal engineer going, I don't believe it. Absolutely not. Not me. Not here. Not. No way. And in fact, last week, I heard a story of somebody who had their most senior staff engineer quit. It says, [1:09:20] And I quote, "I do not believe in AI." [1:09:23] I will not work in a place that does this. [1:09:25] So what is your appeal, sort of engineer to engineer, of why to invest in this? Why you think it's the way that engineer organizations are moving and how you kind of come through [1:09:36] to meet skeptics where they are. [1:09:39] and hopefully see things a little bit more from where kind of Intercom is approaching them. [1:09:44] I mentioned that Intercon kind of had it on easy mode. We didn't have to convince leadership that there's something to this AI stuff. Like, we were pretty much-- had decided the direction of the company the weekend that ChatGPT came out. So we already had this-- [1:09:59] expectation that this will be transformative across many parts of our work, including all of building products and engineering. We were just kind of mostly annoyed about how long it took. But I think [1:10:11] For sure, it does need strong advocates and you need [1:10:14] to push... [1:10:15] boundaries, like one of the biggest things that I've been able to do successfully was kind of push through the barrier of like, should we let an agent connect to Snowflake? Like what, like, and there's all these things can go wrong. Or should we let our agents run real production code in our Rails console over API? And the easiest thing,

[1:10:36] thing to answer there is like, well, you know, I'm not sure. Or like this, this is risky, or we, we should think about this, but we've been largely pushing through it. And now like not recklessly, like we've lots of good controls and we're a mature business. And we have like, I've been on our security team, but definitely not trying to do anything too wild. But there's still, even then I have apprehension just like, is this, I think, I think we should do this, but it seems weird or [1:11:06] myself permission. And then I realize if I have to give myself permission, there's loads of people out there who just need me permission. And honestly, one of the biggest things I do at Intercom is just telling people they can do things. There's a pre-AI and post-AI, or telling them, like, look, whatever you do, just blame me if it all goes wrong. And I'm just going to [1:11:26] And I guess maybe we can blame Claude now, but ultimately it's that permission and [1:11:31] just like there's a level of ambition which comes from as well as like, if you're out there saying, "I'm not sure if AI is going to [1:11:39] take or have a big role to play in all of our work. And you keep on saying that, that kind of will permeate through the culture and people say that. But if you're very clear, you're saying that like, look, all work is going to be agent first. [1:11:52] like at some stage in the near future. And so we're going to [1:11:56] figure out the path there. And so we're going to break down every barrier as we come across them. And look, it's your job, it's my job. And if anything is wrong, blame me. Like that's largely been how I've been approaching it, but not just me. Like this has been a very large collective effort. But giving that kind of permission thing, but also the kind of like freedom to like explore or push things or whatever, it's kind of necessary. And look,

[1:12:18] Mm-hmm. [1:12:19] It might be a less stressful way to go about it to just take a nap for a few years and come back, and then when all the problems have been solved, and we've got these perfect agents running amok in our environments, then that would avoid some of this. [1:12:49] just has to be on that like enablement and giving people space to go deep on the work, enjoy it and have that moment where things click and you start realizing like, oh my god, this is something that [1:13:01] Transform how much I can get to them. Say it again for the people in the back. I love I was like, oh, my gosh, I love this so much. [1:13:09] You know, it is absolutely those two things, which is like give permission. [1:13:15] you you can please just go please by all means go ahead designer hit me with the pr no one's gonna get mad at you like go ahead and then the second thing of just [1:13:25] accountability can roll to the top and not in a scary way. Let's not do irresponsible things. [1:13:31] But, you know, we've seen a couple incidents in the past months, some big ones. And what you see is CEOs or big leaders coming out and saying, like, the team's shipping and we want to keep shipping and we're going to be careful with our customer data and we care for the customer experience. And stuff happens. We've learned from it.

[1:13:51] It's ultimately on me. I'm going to call the customers and we're going to we're going to move on and deliver great innovation for you. [1:13:57] And you know what I tell people to, you know, to get them over that hump, which is like, you really got to know what your existential problem is. And I love what you said is the second that ChatGPT came out. [1:14:07] Intercom changed. [1:14:09] Because that is an existential problem. [1:14:11] Who writes the code in your code base? Agents or humans? Not an existential problem. Like, will you be fundamentally disrupted by a new technology? That is the real problem in your business. So I always tell people, like, [1:14:23] Let's differentiate the real problems in our business from problems that we can tolerate. [1:14:28] And then go use the problems we can tolerate to move fast. And so it sounds like you have a really good call. I mean, I think... [1:14:35] at the end of the day, [1:14:37] the results speak for themselves. And again, you all are not asking me to say this. [1:14:41] Intercom has meant the moment. [1:14:43] You went all in. [1:14:44] on ai assisted you know customer support and experience you're now building models and so it's not just a one and done [1:14:51] chat GPT is here, we need to change how our product works or AI assistive coding's here, so we need to change how our engineering team works. It's [1:14:59] You know, models are going to be how people differentiate. We need to go there. CLIs are going to be how people use products. We need to go there. And so I think this sort of like fearlessness and what I would suspect is like just a fun, nice, high trust culture. Good people. [1:15:14] You actually see the business results on the other side. So I'm going to hype you up. I see a lot of teams. I see a lot of leaders. And I think people can take a lot of inspiration from this. But let's uninspire them really quickly before I get you out of here, which is my last question, which is when...

[1:15:30] Finn takes 15 solid minutes on a live podcast to do a very basic task that you know it can do. [1:15:37] Or not Finn, WinClaw code. Yep. [1:15:40] What do you do? Do you yell? [1:15:42] Are you a yeller? [1:15:43] What does your meta-analysis on this internal dashboard say? [1:15:47] The human needs to improve on. [1:15:50] I do lapse into giving clog codes, like... [1:15:54] just like smiley faces or unhappy faces or, you know, not over the top. I certainly haven't cursed at it. [1:16:02] Very polite. That's kind of not my spoil. But I do like the odd kind of, like... [1:16:07] at a boy kind of smiley face. And I don't know if it knows like that I'm deeply thinking about this and like these little subtle kind of hints or whatever. But yeah, no, I think like professional with a few emojis is my style with Claude. And, you know, hopefully that'll come back to me someday with an emoji. Same. I waste the tokens on telling it it did a good job. I somehow in my mind, I'm like, that's going into into its own sense of itself. [1:16:33] And it's going to know what good looks like. So I am there. I am there with you. [1:16:38] All right, Brian, this has been... [1:16:40] one of my favorites y'all if you have gotten to the end there is so much alpha in this episode i cannot believe it this is a cheat code to winning friends and influencing sass um [1:16:52] through AI engineering. [1:16:54] Brian, where can we find you and how can we be helpful? [1:16:57] I can be found on the internet as a nice vanity URL, which is brian.scanlon.ie. And I got a few links here to some other talks and some other writing and different bits and bobs. As you can tell, I'm not a designer. I asked Claude to design this as if I was a Unix systems administrator writing a little web page, and it kind of shows. I'm active on xTwitter, brian__scanlon. I'm on LinkedIn, ScanlonBee or something like that. I think I'm the most famous Brian Scanlon on the internet. So generally, Brian Scanlon, that tends to work.

[1:17:27] And I tend to be active and like showing up to different conferences and just [1:17:34] like getting good word out about what we do at Intercom, mostly these days AI, but I've also given lots of talks about many other different topics. And yeah, I'm also a big believer in just saying yes to a lot of things. So if you look me up, you've got a good idea, you want to get in touch, you want to run stuff past me or whatever, chances are I'll say yes. And we can, I'll just keep on doing this until things break and then I start saying no. But I'm still not there yet, so bring it on. [1:17:59] Great. So search for Brian and ask him to do something for you. That's it. Well, thank you. So thank you truly for sharing all this information. People are going to get [1:18:07] Tons of value out of this is going to be a hit for sure. And I just really appreciate you joining Howie AI. [1:18:14] Of course, this is so much fun. [1:18:25] You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.