Agents moved where the work happens (and using MCP to find it again) | Slack’s Jaime DeLanghe

0:00Welcome back to Dev Interrupted, brought to you by Linear B. Before we start, my guest today is Jaime DeLang, Slack's chief product officer. And the idea that stuck with me most from our conversation was how the real org chart lives in Slack. You learn how your company works by watching where the work lands. And that's the whole reason Linear B sponsors this show, because the same thing is true of your code. Once AI helps developers write faster, the honest test isn't code gen, it's what happens downstream in review, testing, and delivery. That's where you find out whether AI is really helping you ship.

0:36Measuring AI's real impact, keeping it governed, seeing what happens downstream, those are three things Jamie and I kept circling back to, and they're exactly what we think about all day at Linear B. Now, on to my conversation with Jamie. me. Today's guest is Jamie DeLange, Executive Vice President and Chief Product Officer at Slack. And as our listeners know, I'm a total Slack nerd. I spend a lot of my time in Slack building agents for my teammates. And I've formed a lot of opinions, especially over the last year, about what my workspace needs and what kind of space it needs to bring to the mix to make AI powerful.

1:16And in the last year, I've really had the privilege of being part, I guess, of the Salesforce cinematic universe and going to places like Dreamforce and TDX and getting really keyed in on how Slack and Salesforce are betting on the transformations they're making with their technology, their bets on where the work is actually happening. And Slack is center stage on that. I've heard a lot of great panels and discussions at both of those events about the direction that Slack is going. So I, too, was like, you know, let's bet on this. Let's build on this. And as part of learning more, you know, we had Curtis Kempel from Slack on our show, who even walked me through how he created some of his Slack agents, which was really exciting to dive into.

2:01And that was around the end of last year. And since then, you know, I've launched a whole fleet of Slack agents for my teammates. I found a lot of success building on the things that Curtis taught me. And I've also been so excited about how Slack is deepening its partnership with protocols and technology that are powering AI for everyone and helping all of us enact gains and find the best standards moving forward. So it's exciting to be part of that story. But centered to all of this is how the Slack, the product, is evolving and changing and meeting the needs of knowledge workers now. And that's what I'm so excited to dig into is some of that high-level thinking because here on Dev Interrupted, we love to get aligned about where all of the ships are going because it's pretty stormy seas right now.

2:47So, Jamie, welcome to Dev Interrupted. We're so excited to have you here.

2:52Jaime DeLanghe:Thank you. I'm so excited to be here. I feel like your intro just spiritually connected us in like a really deep way. I'm so excited to talk to you about everything that we're doing. Amazing. Well, now that we're best friends, we're going to dive into a whole chat today about how Slack is transforming. But obviously, the opportunities that you've been exploring as a product leader and what you see the AI, the powered world, the Slack world that we're operating in changing into. So do you want to start with what your vision is and the opportunities you see? Yeah. So I feel like I'm a Slack CPO now.

3:32Jaime DeLanghe:I've been here for about eight and a half years. I haven't been the CPO the whole time. I'm more recently the CPO. But when I think about the future of Slack, as somebody who's just been here for so long, it's impossible for me to not think about Slack's past and the things that have always made Slack Slack. Made it the place where people love to work. Made it the, you know, only piece of SaaS that, you know, people like really like fanboy over. and I want that to continue to be true as we all evolve how we're working. And so the things that come to my mind in terms of like, what does Slack always been and needs to keep being and maybe even get better at is one, it's the place that people want to work in.

4:14Jaime DeLanghe:It is human, it's friendly, it's fun. It works the way that people work. It's not trying to like put some rigid structure on top of how you get stuff done in your day, but it's gonna be helpful. It's gonna be there just in time to help you get a little bit organized, maybe help you prioritize your work, help you and your team stay on task. Slack has also always been a platform or for almost the whole time Slack has been around, it's been a platform. And we've had this strong belief that the more work you bring into Slack, the better your workers' lives are going to be. They're not going to have to go into every single SaaS application to get stuff done.

4:55Jaime DeLanghe:And the more people can look at information together in a channel or look at content together in a channel, the more aligned they're going to be, the more they'll be able to move. And then finally, all of this needs to be open by default. It all needs to be searchable. It all needs to be findable. So when a new hire starts on day one, they don't have an empty inbox. They have like a whole bevy of channels to like dig back into, and they can basically get up to speed super fast because they're just jumping into the conversation. But they're not missing the whole rest of the history. They have everything.

5:26So all these things have always been what Slack is,

5:29Jaime DeLanghe:but now we got agents, right? Like we have LLMs that can come into the mix. And a thing about me is that when I started at Slack, I worked in search and machine learning. So I had all of these like highfalutin ideas about how you're going to turn all of that information in Slack and turn it into knowledge. And we were going to make like Slack the smartest application is going to be super assistive. And we can actually do that now. Like we can do that. Slack can do that itself with Slackbot. But more than that, we can be an engine that kind of helps every developer, no matter who they are, do that for themselves, for their companies.

6:05Jaime DeLanghe:And so when I think about the future of Slack, it's really extending that original vision of being the place where all of your people, your applications, your information come together in channels to keep everybody aligned and moving forward in the right direction. and extending that to agents and making sure that Slack is the best place for humans and agents to collaborate and making sure that that human and agent collaboration stays human at its core and that we have the same kind of flexibility to work in the ways that make sense to us, that we're not being pulled into like isolated AI chambers where we can only work with robots and then we have human spaces where we can work with human spaces.

6:43Jaime DeLanghe:I think these spaces need to blend a lot. And with agents, it also means we can pull in a lot more work. Like we used to say that Slack is like, you can pull in the thin work that's just really fast, like web forms, web hooks. Turns out those were super powerful, right? But now you can do so much more. You can paint graphs inside of Slack. You could maybe even like, you know, review code inside of Slack. You could completely build your whole website from inside of Slack. And that's what agents allow for. And I think as I think about what are we going to build toward, it's continuing to build that environment and that ecosystem that really makes the interaction patterns powerful and delightful and thinking always about the team and how the team gets work done, not just about an individual person.

7:32Jaime DeLanghe:So I think Slack needs to evolve to make humans and agents work together super well, make it the best place to develop your agents, deploy your agents. Slackbot can be sort of a super orchestrator over all of that in so many ways. And then we just need to keep making the experience dead simple and delightful so that people want to stay there. Exactly. Dead simple and delightful. And the way that it has to transform to meet the individual needs of the worker and what's important to them and the organization in general. What's really exciting is, you know, this vision that you see for Slack is this like context world, this equal plane operator where agents and humans can work together and have the views and outputs that they need to get their work done.

8:18Really what it exposes is that much like engineering and the ways that we solve code generation and delivery at scale with AI agents, almost that same thing is happening with our knowledge work. And the sooner and more clearly that we have understandings about where are those loops within our org, what are the windows we need to have into these loops? And these become channels. You're like punching holes into your organization to see where these agents are working. And Slack becomes where you're peering in to do that work. And then, you know, there's some stuff in there that we're going to touch on a little later when we talk about MCP and protocols and stuff.

8:56Because then obviously you can bring the whole world into Slack now because everything can get bridged into AI conversations. AI conversations can take advantage of MCP, which can have apps embedded in them now. And we can turn natural language and conversations that are happening into live rendered things. Like you said, you can drive any kind of process that you need from there. So then we arrive at a responsibility for a lot of leaders, both non-technical and technical leaders, product leaders, engineering leaders, everyone alike, to think about how are we going to engineer the harness of our context world, of our Slack world, of our communication world.

9:37Because now that we've acknowledged that, oh, this too is a bunch of loops and a bunch of checks and gates and guards and contacts that goes in and out, how are we going to build it? And what I've found is that Slack becomes the building block to not only start building it, but express it really quickly, which is its superpower, I think. And that's like it's in a lane of its own in terms of where you can take it.

10:00Jaime DeLanghe:Totally. I mean, I think like Stuart Butterfield used to call channels. they were digital containers for arbitrary objects, which sounds like, you know, that's kind of nonsense. But exactly that, right? Like you can, one of the toughest things when you're building an agent is how do you figure out the context layer? How do you like build up your knowledge graph that's going to power this thing? How do you keep it up to date? And one of the coolest things about Slack is it has just enough structure for you to be able to like point at like a project or point at a topic, but not so much structure that you're having to stitch together a bunch of different things.

10:40Jaime DeLanghe:And the conversation itself actually ends up being like the richest context for you to be able to action off of any of like, I don't know, the meatier things like a slide deck or even like a PR. Like the more you know about what happened around it, the smarter your agents can be. Yeah, you're like scattering all these artifacts And all the roads, something needs to lead back to Slack. It all comes down to good, healthy reporting, information going into Slack, communications happening in Slack, and correctly labeled semantic channels that are open and not siloed. You know, good communication practices we should all be having as organizations in general.

11:19Because it turns out AI amplifies the good and the bad. Yeah, exactly. And so before we dive into some of the more exciting technology and product changes that are powering the ability to do all of that, I do want to ask because I'm curious from a practitioner standpoint. The barrier that gets crossed from when you're doing the single-player AI experience to the multiplayer one experience, it becomes profoundly more complex in terms of how you deal with the interactions. that's something that I've learned just building with them but also too just as like someone interested in the space two years ago I gave a talk at B-Sides Barcelona about multiplayer AI and how you could use AI chat conversations to basically hack your co-workers context like this is an understanding that we need to have as folks that are operating in the space together so how do you think about as a product leader about securing that kind of space and making sure too that it's like going to scale for the needs of all of your users that have a huge range of sensitivities for their topics.

12:22Jaime DeLanghe:Yeah, I think it's really interesting. It's a topic that we're continuously grappling with. I think one of the benefits we have here is channels. And so channels are not only a digital container for arbitrary objects, but they're also an ACL basically. And so you can kind of make a pretty good assumption that whatever is pushed into the channel, if an agent has access to that and the people have access to that, that's a fine, we're all happy. That's a happy universe. It gets a little bit more challenging when that agent gets access to tools. And it's not just looking at the content that you put inside of Slack, but it's looking into some third-party system.

13:04Jaime DeLanghe:And where we are at right now with that is really allowing for configurability for whatever use case makes sense. So in some cases, it might make sense to have an agent have fully its own identity and have that identity, like it had its own account in whatever tools it's having access to. It may also make sense for it to be an OAuth where it's passing through your identity and whoever's talking to the agent, now the agent's acting on that person's behalf. It may make sense to throw up a modal or a modal, but like a radio buttons or something to be like, do you want to do this as yourself or do you want me to do it as me?

13:44Jaime DeLanghe:Like to just like explicitly ask the user, I think what has become imminently clear to me as we dig into this more is it's really difficult to make a rule that fits everything. Like everybody's going to have, like you said, a different level of sensitivity. Everyone is going to have a different use case. and configurability. And even just, even that idea of like on the fly configurability, I think could be really powerful so that we're fitting the configuration to the use case and letting the people who are closest to the work actually make that decision. And I think then it's just about helping people understand the implications of those decisions, whether it's on the developer side or the admin side or the end user side, because there's trade-offs.

14:32Jaime DeLanghe:There's no perfect answer for what that looks like until, you know, the agents get so smart that they know what's secret and what's not. But I don't trust them yet. I don't trust them quite yet. I don't trust them yet. You're hitting on exactly kind of how, when we've talked with other leaders about how they think about the security planes, the world within their space, how they think about it too. It has to be deeply configurable just as much as it has to, you know, the needs of your end users are on a wide range of sensitivity. your configs need to expose to that and meet to that level. But another thing too that we learned recently, we had Tanya Janka, who was OWASP Hacker of the Year last year, and she helped put together the OWASP Top 10 vulnerabilities for 2026.

15:17And the top one was like vibe coding and developers themselves. Like people have become the biggest vulnerability, which is why I think it's important for people to double down on using places like Slack and using like user management and control plane surfaces, like what's offered in like the supporting Salesforce universe to really understand and identify all of the movers within your system and what they have access to and being able to scale it up and down with the levels of like sensitivity that you need. And, you know, I think it's really easy to like, you know, roll your eyes over some people like just to not necessarily want to double down on that.

15:54But it's like, ultimately, if like your workflows are just logs on a machine somewhere, they're like so sensitive in terms of the amount of work going in and the outputs. Whereas if it's like siloed away in a specific kind of like channel, a chamber for that particular purpose, you can understand and audit its usage. You can control it a little bit better. And ultimately that's like the kind of control plane that I think knowledge workers need.

16:19Jaime DeLanghe:We've also been, I think about it at like a administrator level. So like, you know, more fine-grained provisioning for agents, being able to say, like, these agents can only be operated by these people, potentially getting to a place where we say these agents can only operate in these channels. This is an allow list for the channels that this agent can be in. That's something that an admin could set. But the thing that I think we haven't yet quite figured out is, like, a lot of the people are not going to be admins. The builder also needs some visibility into this and as the bar, like, the barrier to entry on a building It's so much lower.

16:57Jaime DeLanghe:Like we need, the builder should be able to do that. And we should make the builder as close to the business owner as possible so that the business owner could also know that. And ideally, maybe they can't build the agent themselves, but they could troubleshoot it. Like they could push it in the right direction and see where it's going wrong. And I think this is, it's something we've got to figure out for any, both agents and and things like skills, anything that you can pass around and is getting leveraged in multiple places. Having that kind of observability so that you can inflect on what you've built, it's really important.

17:34Jaime DeLanghe:And also figuring out how to make that understandable because this event is not just software. It's not like there's a bug. I wrote the if statement inverted or something, And now the whole thing went kerflooey. It's like much more nuanced and helping people, maybe even with other agents, figure out how to inflect on their agent to make it better. I think, I don't know what it is yet, but I feel like there's definitely a role Slack can play in making that world work better. Your teams adopted AI and now leadership wants proof it's driving real business outcomes. Linear B is the engineering productivity platform that measures AI impact at the pull request level.

18:19So you can see exactly how AI is shaping speed, quality, and efficiency across every team. Linear B helps engineering leaders reinvest productivity gains from AI back into their engineering organization. Stop guessing at your AI ROI and start proving it. See how at LinearB.io. Yeah, I'm even thinking now about how Slack can help organization leaders identify and highlight and elevate these AI unlocks that do happen, often in a really siloed basis on like a team-by-team level. And then like what you said as well, you get a level of skill sharing that becomes really important. Slack becomes a really great, almost like skills ecosystem with all different ways that it can plug in.

19:01So, you know, skills themselves, an open standard, one that AI users around the world are just like really strongly embracing. And another one as well as MCP. And MCP is, of course, like been a little bit of like a news hype darling child. It has been the coolest thing ever. It has been dead. It has been brought back to life. It has been really mislabeled, relabeled. Bets have been taken on and off the table. And along the way, I think MCPs totally had an identity crisis, but the one thing along the way that it did better than anything else was distribution. It had a lot of really good stuff, really cleanly in a lot of people's hands in a really scalable way, which, hey, guess what?

19:41It's perfect for AI rollouts, AI enablements. Now folks are embracing skills over MCP, has a new working group with the Linux Foundation. I'm like, can that spec please just hit the MCP protocol already? so we're all locked in right on how these protocols are building for the future so you know how is slack also playing in that space and you're bringing things like mcp into slack what is it what has it been like to watch those technologies unfold but then also to participate as like a

20:07Jaime DeLanghe:builder a contributor well i think one of the things that's very um we have the great fortune at slack of having such a strong developer community and such a strong partner community So I think for pretty much anybody who's really building agents in the ecosystem is thinking about building them in Slack. Our partners are incredible. The individual developers at our companies are incredible. And they're all telling us what they need from us. And the move to MCP was pushed in a lot of ways by the community telling us that it's what they needed. That's when the MCP server came out because we saw the community trying to do a lot of things with our APIs that they weren't designed for that felt to us unsecure and uncomfortable.

20:57Jaime DeLanghe:And we were like, we got to figure out how to meet this demand, but we've got to do it safely. And that was our first move to MCP on the server side. On the client side, we really started to play with things internally. We built Slackbot ourselves. We started connecting our own MCP to it. I think originally we had this idea that we would only have a limited set of partners for MCP because we weren't totally sure on what the level of adoption was going to be. And we weren't totally sure on what the level of sophistication of that adoption was going to be. Like, are people going to just publish junk to this and degrade the experience of the agent to such a degree that, like, it's not really usable anymore?

21:43Jaime DeLanghe:And we kind of solved for that by allowing our own MCP ecosystem to grow internally. And, like, we're not, you know, we're not Luddites. We understand how to build technology. but we did have MCPs of varying quality and that allowed us to understand that like, this is, we're gonna be okay with this. We can do it. We can have good tool selection. We can have good logic around the tool selection. And leveraging MCP is gonna give us a lot more power than trying to roll our own tools for every single thing that we might want it to do or trying to figure out a more robust like agent to agent protocol.

22:18Jaime DeLanghe:I think this will get us there a lot faster and provide a lot more flexibility. I also say like highest level, I want anybody who is thinking about building for agents to be able to build for Slack. And I also want to thank anybody who's thinking about building agents to be able to easily put those things into Slack. And that means that we have to have open standards, not Slack-specific standards. So, you know, I don't think MCP is going to solve everything for everybody. And I am, like you, very interested in skills over MCP. I think we should get that as soon as humanly possible. Because it actually just makes skills and MCP go together like peanut butter and jelly.

23:00Jaime DeLanghe:Okay, great. You need both of them. We're all literally the same page. Yeah. So I agree with that. I also think, like, you know, there are places inside of Slack where we're playing around with 8AA. Salesforce is betting on ADA. We're testing a bunch of different things because we don't know where this whole thing is going to go. So I think being tinkerers is really helpful here. And then having our ear to the thousands of other tinkerers out there who are trying to get their stuff inside of Slack, I think our ecosystem pulls us in the right direction a lot of the time. Yeah. So here's the playbook that I'm hearing then is that you're standardizing early because it's less expensive to standardize early, but obviously keeping a very open ear and open conversation with all of the tinkers abound internally and externally to really go through what works and what doesn't work.

23:58And that requires obviously like a good engineering practice, a good reflection practice, a good ability to like highlight things that are working or figure out like why is this not working. But also to when you bring everyone together to work on that unified problem and you give it a label, then it gives you an opportunity to to be such an early influence on where that technology can go. And MCP is one that I think anybody who's building in the spaces that we are, where we're building things that are connectors and highways of information for other tools, MCP is absolutely going to be critical for business success, for product success, just for the technology to be safe and scalable.

24:39So acknowledging that and knowing that MCP has to be center stage, it's important to partner up with those protocols early. And this is something that I learned and got a really good deep dive from when we had Tyson Singer. He's from Spotify backstage and really the kind of person behind the IDP portal. And what's always been so fascinating to me about that product is that that's not what Spotify is out to do. Why do they have this first-in-class IDP and they set the standard and everyone just uses backstage because it's awesome, right? But it's like when I really dug into his story and he told me how, you know, when they were early and figuring out what their IDP was going to look like, it was more strategically and financially and long term smart for them to just like set the standard, make the standard, go with the standard because it was so early in the play.

25:30Because switching later would be so expensive. but he was saying too in that same conversation that he was like but if there would have been something really great to latch on to that solve that problem like that's what we would have built around because that's what we're optimizing for so that's why mcp is now i see that for a lot of big orgs and enterprises they're building mcp like things into their product and hedging their bets about where things might go but ultimately we're all still just like one prompt away from doing like the full changeover when everyone else gets on board with uh because that's how we engineer these days when everyone gets on board with where things have got to go.