Multi-agent orchestration in Slack | Saleforce's Kurtis Kemple

**Andrew** (0:05)
Today, I am thrilled to welcome our guest, Kurtis Kemple, the Senior Director of DevRel at Slack. Kurtis, welcome to Dev Interrupted.

**Kurtis Kemple** (0:15)
Thank you so much for having me. It is a pleasure to be here.

**Andrew** (0:18)
We're really excited to have you here. You and I, we met at Dreamforce last year, and when we met, I knew I had to have you on the show to pick your brain, because we chatted for a while about some pretty cool concepts. One of them really stuck with me around the future of work. I really want to dive into that with you today, because when we chatted at Dreamforce, you shared this idea that really stuck with me about how Slack is evolving from a place where work is discussed to where the work is actually done, as if those words are starting to move into action. I think that's really interesting to explore. It makes me think of supporting practices in the code world, like DevOps. Maybe we're entering a world where you get something like ChatOps. This is part of the future of work that Slack is taking us and everyone who uses Slack, which is a lot of folks, into the future. I want to dive into that vision with you and talk about how those core problems have evolved.
What do you think about that premise? Do you want to dive into that today?

**Kurtis Kemple** (1:20)
I absolutely do.
Before we hop directly into that future, I just want to take one second to talk about the past and how we got here, because Salesforce has really created the push for Agentforce and Agent Interaction into Slack. That was our real first approach, right? We prompted in there. We learned a lot. Through that process, we started to understand and develop, what does it take to support that type of experience, getting agents directly in, first through there, but now through anywhere, right? Like third-party directly into the Slack platform or first-party customers building their own agents and integrating. So we just hit that perfect storm or shelling point, if you will, right? Yeah. Essentially, it made us really just stop for a second and put on a beginner's mindset and say, what is a platform that supports any agentic workflow?
But does it in a way that is structured, consistent, grounded? That is a very difficult tension to think through. So I just want to preface that, and we've been working with a lot of customers to figure this story out, right? Like Anthropic and Versel have been at the forefront of this. We've got all kinds of companies really just helping us grapple it, another one that stands to mind, just tons of these across different industries, all noticing and saying like, hey, we can deploy AI here because we've got collaborative environments. We've got context, which is what we're going to talk about here. And so yeah, so sorry, just the main intro, I just wanted to bring us because that's how we started thinking about the future, right? Like, look at where all these things are heading. There are some similarities, some things that are overlapping. And when we think about truly having humans and agents working together and collaborating, like, what does that look like in reality, right? Not even just at the code level, but literally handing off at the interaction point.

**Andrew** (3:35)
I love how you framed it. I'm really excited to dive into this because you're right. I mean, Slack becomes the place where all of that context lives. And that context is messy. It's the real communications between real people getting their work done. It's not this neat, orderly, structured data that can flow in and out of systems. And so it creates this perfect intersection between the systems we're building, to be more productive, and how and where the work is getting discussed. And it's exciting to think out how all of these other companies to see the opportunities with their conversations and want to tap into that to make their own work better.
And really, it comes down to this context, right? It's because context has evolved now into being a first class citizen of the AI world. Before, we were all about prompt, some prompt engineering, and then it evolved into context and context engineering. And I can't think of a better source of context for a lot of the things that happen at work than maybe some Slack channels. So can you expand a little bit on this context gap and how it really is needed to help models perform and meet companies where they want to use it?

**Kurtis Kemple** (4:40)
Yeah, absolutely. So I'm going to walk you through super quick something that I refer to as leaky prompts, right? When you only own half of the experience, meaning that I can't control what a user prompts, right? And they might start off with a very perfect prompt with what they're trying to accomplish. But literally proven through science, like any conversation, whether that's with something digital, another person, a group of people, will actually slip into chaos unless it is actually managed, like triaged, right? And we see this actually, you do this right now. When you are interviewing people and you got engaging conversation and we're chatting, that takes effort from you and energy. You are literally putting in a ton of work to ensure that we have this very good fruitful conversation that stays on track and has important insights and talking points. So, you know, that work is also required when you're engaging with an LLM, surprisingly enough, right? But the issue is, is we can't control how somebody else is doing. And so it puts us in a place where the only way that we can have the best chance of ensuring that that intent is in alignment, we're staying on task to their goal, is that the context, the representation of what we give to the LLM on the user's behalf is as best a representation of what we can think they're trying to do. You're almost adding like a second order need of understanding, all right? And it's like, you have to understand how the user is going to interact with the LLM and ensure that you can just provide the right context. I like to think of it more as information architecture at this point. And if you can do it well enough, it makes it a lot harder to have those conversations get off track and that misalignment on a tent. It makes a difference. And like you said, Slack is a wonderful home for that context. We've got threads and channels and messages. And that's where I see the secrets all set.