From Models to Harnesses—The Architecture of the Agentic Enterprise

Julia Nimchinski:

And we’re alive. Welcome to the Agentic Harness Summit. Over the next few days, we’ll focus on how companies are turning business processes into business harnesses and protecting their alpha and IP, building adaptive GTM models around proprietary context, workflows, governance, and guardrails. We’ve got an exceptional lineup of CXOs, founders, and analysts, and we are kicking this day off with Martin Kihn, SVP of Strategy, Marketing Cloud at Salesforce, and Russell Scherwin, Facilitator at Force Management. Former IBM CMO, Managing Director at B2B Tracks, and MBA professor at UGA. Welcome to the show, welcome back, Russell. Welcome, Martin. How are you doing?

Russell Scherwin:

I’m good.

Martin Kihn:

Yeah.

Julia Nimchinski:

Amazing. Russell, take it away.

Russell Scherwin:

Awesome. Well, so excited to get this started. You know, before… I’ve always, found that with any message, it’s not what you say, but sometimes who says it. So, Marty, why don’t you, give… give this audience a hello and some context for you?

Martin Kihn:

Yeah, so I’m Marty Kihn, as was mentioned. Thank you for, it’s really an honor to be kicking off the Agentic Harness Summit. Harness is a very… I’m sure we’ll get to it, Russell, but, my role, as was mentioned, I’m a SVP of Product Strategy for, Salesforce Marketing Cloud. I work on the Data 360 with CDP, and also Agentforce, which is Salesforce’s version of the Harness, so…

Russell Scherwin:

Fantastic. And hey, I’m Russell Scherwin. Julia introduced me far better than I will. Yeah, I’ve been a CRO and a CMO. Currently, I help organizations infuse their strategy into every market-facing conversation. And this weekend, I was in the process of building out my website with Claude and, Marty, it took me probably about some weekend time to do what would have taken me months back when I was a developer years ago. But hey, that’s, you know, first half 2026 stuff, where we’re into newer and better things, and when I talk about Harnesses here. Now, I liken giving agents, you know, you know, an intelligent agent the keys to run my business operations, just like giving your key… your kids the keys to a Ferrari, because they passed their road test.

But before we get into talking about harnesses. Let’s talk about what the heck a harness is, because otherwise we’ll be in a game of buzzword bingo. So, Marty, start with defining what a harness is for the audience, to kick this off, right?

Martin Kihn:

That’s a good question for the Agentic Harness Summit. Let’s ground ourselves in… I think, there was a documentary on Netflix, some of you probably saw it, about a guy who climbed El Capitan, the bare rock phase, without a harness. He isn’t crazy, but it’s good. If you’re the best climber in the world, go ahead and do that. The rest of us need a harness. I mean, I think of it this way, an LLM, Large Language Model. does one thing. It just predicts the next word, given everything that came before. So whatever’s in the prompt, it’ll predict.

That’s all it could do, but there… the harness is everything that’s around it, so you can put it to work in your enterprise. You have to ground it in the context of your enterprise. You need to do identity resolution. It has to have memory, you know. What happened in the last conversation? What did the customer do last month? There needs to be governance. Who has access to what data and what context. There needs to be, orchestration, so that’s basically triggering actions, and then action is the key word here as well. putting AI to work, giving it tools, allowing it to do things like send an email, send a query, update a record.

All that stuff can’t be done by an LLM on its own. So the harness is really the infrastructure around. The analogy would be, like, to a car. The LLM is like the engine. And the harness is everything else. The wheels, the steering wheel. Yep.

Russell Scherwin:

If you just let the engine run, without steering it.

Martin Kihn:

Not gonna go anywhere.

Russell Scherwin:

Excellent.

Martin Kihn:

Yep. Yeah, you’re stuck in neutral. Yeah.

Russell Scherwin:

Fair. Okay, so the… so basically, and I… if it’s okay, I’m gonna go with the analogy that stuck in my head, which is, you say, the LLM is… it’s the dumbest power source. It knows how to speak English, and it has some level of intelligence. It’s kind of the power source. But what you’re talking about with the harness is harnessing that power, and you talk about identity, you talk about memory, you talk about orchestration, and then you’re putting it… you talk about putting into action in use, so your business applications can use that power. Can you talk about… let’s go from the, here’s what it is, to how have you seen it working in the real world?

I mean, as an SVP of Strategy for Marketing Cloud, you’re seeing I imagine some incredible use cases. Give us… give us a flavor for one of the two use cases of how organizations are harnessing this power, let’s say.

Martin Kihn:

Yeah, sure. I mean, the two most common now, we have… Agentforce is like an agent factory, you build agents, deploy them, monitor them. The two most common uses of that, so this is sort of our internal data, if you will, are, number one, customer service, so routine customer service inquiries. And if you think about the context customer service in retail, apparently, you know, it’s quite predictable, people asking the same three questions, you know, where’s my item? When are your hours? How do I return this thing? So, it can be identified, it can be automated, but you need a harness around the LLM, because otherwise the LLM is just providing… it can answer questions, but it won’t remember the context of the conversation in the past, it won’t be able to look up your customer record.

Without the harness, look up your customer record to figure out, you know, what your purchases are. It can’t go into the database to find out what the hours are, all that stuff. So that’s required. The other major use case, other than customer service, is in lead nurturing in the B2B context. So, people come to a website, they, you know, watch a webinar, they download a white, whatever, they become a lead. So you have their email, but most of them are just ignored. You know, they’re scored, and they’re like, oh, they don’t meet our threshold. If you can automate every single one of those leads using an agent, so it’s not a person doing it, it’s an agent, you can nurture the leads, and some of them will turn into customers.

So you don’t have to, like, you know, neglect The long tail that you couldn’t get to before, just because you didn’t have enough people.

Russell Scherwin:

Cool. Well, let’s play with that first use case, because we can all put ourselves in the shoes of someone who needs to call up a retailer and say, hey, I just bought this a week ago, where the heck is my stuff? We’ve all… we’ve all lived that use case. Let’s define three worlds, and I want you to walk… and if you have a customer real-world example, even better. World one. Let’s say before AI, which I think is straightforward, but how did, let’s say, a retail call center handle that before? Then let’s go to a world of AI agents. without the harness.

And then let’s talk about what the harness brings to the table that AI agents running amok the engine into the wall were not.

Martin Kihn:

Well, in the old world, we remember it, it was, like, last year, but basically, you would call up, there might be… there’s an IVR, so there’s kind of a primitive version of AI, I guess, which would be a phone tree. And you were handed off here and there, and then you got into a queue, and depending on, probably on your status, or… When you called. you would get connected to a person. A lot of times, many of those calls took too long, people gave up, whatever, or they were referred to the website, because they wanted to handle it in an automated fashion.

So that was the… that was the previous version. A lot of calls were neglected, or a long wait time, bad customer experience. Many of them were handled. Then, with an LLM alone, that could be, like, for instance, like, I call up, the call center, and I ask a question. Where’s my order? The LLM can’t tell me. It can’t look in the database, so it can say, oh, let me refer you to a person, because they have access to these databases. You could ask it a question about a product, like, do you have this in… not, do you have this?

Does this item come in brown? And they may have access to some kind of product information like that if you fed it to the, to the LLM, but it can’t do any kind of follow-up, and it can’t… it can’t resolve the issue, which is why you called in the first place. With the harness, you can call up. you’re speaking to an AI agent, and they identify themselves as such. You’re perfectly comfortable doing that. You’re like, where’s my item? They can look it up, so they have access to the database. They’ll give you the information. You can make a change.

You’re like, well, I won’t be home that day, could you send it the next day? They are capable, autonomously, in many cases, of going into the database, changing the date, and the whole thing is resolved very, very quickly. It’s very efficiently, it’s efficient on both sides, and it didn’t require any human intervention, and it didn’t require a long wait time.

Russell Scherwin:

So, so what I’m hearing is the harness becomes the doorway from the raw power, the raw natural language speaking power of an agent to The, let’s say, the deterministic capabilities that a company’s enterprise systems what… Is that… is that fair?

Martin Kihn:

Yeah, I mean, I think I, It’s important to mention that why are people talking about harnesses so much, and, lately, last 6 months or so? It’s because it… they came from OpenAI, and then Claude backed it up, but basically, they… they made the observation that if you… you can use the same model But if you use it with a different harness, you can have much better results, like, two times better. And you haven’t changed the model, like, it’s the same exact LLM, it’s just changing and improving the harness around it. So that kind of harness architecture.

And then the other thing is, we used to talk about prompt engineering, improving the prompt. Improving the prompt makes a very minor, difference compared to improving the quality of your harness. So I think it’s, you know, it’s… Hot for a reason. So…

Russell Scherwin:

So, for these harnesses, I mean, harnesses are this middle layer between the raw power of an LLM and what an enterprise wants to do. Who owns the harness within an organization, and… or the harnesses, let’s say?

Martin Kihn:

I mean, it’s a good question. The, I mentioned Agentforce, Salesforce. I actually wrote a book about Agentforce. I do a quick plug for that book, I recommend it. highly, excellent, really good, good primer. But, the Agentforce, the thesis behind Agentforce is that there are elements of the harness that you, the enterprise, will want to own. Like, you want to own your own data, you own your own customer data. There are other elements, like setting permissions in an application that you’re, you’re okay, you need some help with. So you need, you need a framework, a way to enter permissions and governance and so on.

And then, things like an orchestration layer. So, there are elements of, like, Agentforce where the customer would own it, you know, their own data, their own workflows, their own business context. And then elements where you’re kind of renting it from the software. There are other ways to approach the harness, where it’s all… it’s all… it’s the same old dilemma in the past, you know, build versus buy, you know, well, coming from IBM. But basically, you can do open source, like DeepSea, open source. Claude Code is a… is a harness, so it is possible to have more control if you have a, you know, highly sophisticated engineering organization, and you want that, and you don’t want to have any kind of external products.

So it’s… there’s a spectrum, I would say.

Russell Scherwin:

And I actually want to play that as a spectrum, because I agree. Like, you’re Salesforce, I come out of IBM years ago. I’m working with a client right now where we’re building an AI SDR, and we’re building the harness, you know. pretty, pretty simply and rudimentarily from the logic of the customer’s messaging and the customer’s buying process. So, a harness can be something that’s very big at scale, very IT-oriented, but at the same time, a harness can be something that a CRO or a CMO puts their hands on the keyboard and simply defines the rules of the operation that they’re trying to orchestrate humans and agents within.

Is that a fair statement as well, Marty?

Martin Kihn:

Yeah, it’s true, yeah, and a big, Agentforce is the only one, but basically these big enterprise platforms, the goal of these things, the reason they’re in the market, is that, enterprises, many enterprises, want to be able to build kind of a factory, so it’s not just one agent, or even just an SDR agent, or… it’s a whole group, you know, a whole world of agents, hundreds of them, thousands of them, and they need to be able to create them, test them, deploy them. decommission them when they’re not working, you know, monitor them very carefully, provide government, all that stuff.

So, that kind of a framework is important, and I keep saying enterprise, but, you know, bigger company context. You want brand safety around everything you’re doing with agents, so there’s a lot of control you want around it. And that points to something like a platform, like a, like a, you know, an enterprise harness. Building your own is absolutely possible. It would be more useful, I think, in more contained contexts, where you’re building, like, a singular agent or just a small group of agents, rather than this kind of factory that goes on and on.

Russell Scherwin:

Tell me a bit more what you mean by a factory of agents, because I think there’s a huge abstraction and a huge idea built within that.

Martin Kihn:

Yeah, I mean, I think the important thing about agents in the enterprise context is, I think, you want to democratize this. You want to make it so that anyone could have… not anyone, but, you know, most people could have a good idea and say, maybe this is an enterprise agent opportunity here. How do I build an agent? How do I keep it safe? There needs to be tools that provide that framework around the building, so creating an agent. You know, AI, when you sit down in front of a blank screen, and you say, alright, build me an agent, most people wouldn’t even know where to start.

But if you have a kind of enterprise Agentic platform, it has templates, it’ll provide… you can actually ask it to start for you. You can ask the LLM to start for you. And then, all the controls around it, brand safety, toxicity detection, making sure that the first-party data doesn’t leak. Ensuring the internal governance, that is, who has permission to access what data. All of those very essential controls that you need in the enterprise context are, built into these kind of, you know, larger enterprise platforms. When I say factory, factory is, like, repeatable, it’s just, you know, you could spin up agents, it’s a framework for building agents, yeah.

Russell Scherwin:

I dig that, because that goes back to a pattern that we’ve seen forever, and probably will remain forever as long as there’s enterprises, and that’s, you know, the concept of centralized control, decentralized execution. You know, all great organizations tend to federate power out, you know, to the endpoints where people are touching the market or touching things. And what I’m hearing you say is, you got the concept of a factory, that way people don’t recreate the wheel. So, if I have an enterprise concept of identity. Clearly, identity, authorization, security have to get baked into any agent, but for VP of Marketing to tell their demand gen or their lead gen person, they’re going to have to build, you know, identity management within an agent from the ground up.

It’s crazy. So it sounds like your factory creates the wheel once, or creates a holistic wheel, and then delegates the ability to modify within parameter down to different functions and geos and parts of the business.

Martin Kihn:

I mean, AI, very powerful, you know, we’re all very impressed, but it needs to be watched. It needs to be supervised, not just because it can hallucinate, we all know that, but basically in order to keep it on the rails, like, you need it to be doing what you want it to be doing. AI is… it’s not deterministic. Like, you literally don’t know what’s gonna come out the other end. That’s its power, but it’s also the weakness. So, a harness is, like, it’s literally like with a horse, you have a harness. Why do you have that?

Well, it keeps it under control. And literally under control. Like I said, brand safety, toxicity detection earlier, grounding it in your customer data, so make sure it’s not making up stuff about your customers, but also monitoring. You watch these agents in action. What are they doing? What are the results? And continually making changes, they can… some of them, you know, can autonomously make changes for themselves, improve their own performance, of course, but continually making changes, and if they’re falling short, you decommission them, you know, you kind of fire your agents. So it’s not that different than having a human workforce, this kind of hybrid human agent workforce.

Where you’re hiring agents, you’re giving them performance reviews, you’re decommissioning them, you’re, you know, putting them on an education plan, or whatever. So, it’s quite similar, actually.

Russell Scherwin:

Marty, how do you envision work happening in the future. I mean, let’s go back to your other, use case of lead nurturing, right? There’s… there’s someone in an organization, let’s call them the CMO’s, you know, operations person, but… but somebody’s gonna build out what a harness looks like that drives lead nurturing. Maybe it’s Salesforce, and then there are people who are going to nurture leads, and there are people who are going to write copy, and then there’s agents who are going to write copy, there’s agents who are going to decide what the next best action is.

What are the roles that you see existing in an Agentic world driven by harnesses? And who’s driving those roles, and how do you see it playing out?

Martin Kihn:

Yeah, I mean, I think the future of work is a very interesting topic. I think the… The thesis in the beginning was that agents would replace people, pretty quickly. It turns out that’s not what’s happening. What’s happening is that people are using… enterprises are using agents to augment the human workforce, so the example I gave before, you have customer service, you have a limited number of teams, you can’t… serve all the people you want to serve, but it’s just because you didn’t… you didn’t have the resource to hire that many people. But you can expand your team, and the same thing with lead nurturing.

You can just… it’s like you have a lot more SDRs than you did in the past. So you can handle the long tail. I think that the future of work is going to be… I look at it this way, try to be optimistic, and I am optimistic about AI, but it’s basically we all get a promotion, so even an entry-level employee will come in, and they will be managing managing a team of agents around a specific area. So they… there will be less human beings involved in specific tools, knowing, like, being an admin on a particular tool, and knowing exactly where to point and click.

And much more around, what do I need to get done here? I’m the email manager. I need to know what’s the point of an email, what is the purpose of an email? And then I have agents, sub-agents, who are doing things like writing the copy, figuring out when to send it, you know, monitoring deliverability, and all that stuff. And I am the supervisor, in a sense, of these agents, so there’s a lot more kind of human watching over the AI in the future. And I think, you know, the roles, they do change, some roles do go away.

New roles are created, you know. Prompt engineering, mentioned that earlier, that was a whole new, kind of. career path. That didn’t exist 3 or 4 years ago, and so that’s why the future works so hard to predict, because these new roles appear that we really… we can’t determine what they are in advance. It has to kind of happen organically. And I think the CMO remains. The CMO becomes the master orchestrator and try to ensure the brand voice and all that stuff.

Russell Scherwin:

For sure, mate. I had a CMO, talk to my class, last semester, and, she indicated that the board hired her and gave her an edict Don’t start hiring people until you’ve built your operation and can see where the agents play, and where the humans play. So I think that’s… That might be where we’re going, although I’ll throw a challenge at you. I tend to agree, right… currently, I tend to agree. Personally, I can supervise agents for almost anything that I’ve done with some level of expertise. But, you know, I teach an MBA class. I’m gonna have students, and I teach even an undergrad class at this point.

We’ve got kids who are coming out of school who don’t have expertise. How will they get to the point where they can supervise the agent, as opposed to the other way around?

Martin Kihn:

I think that the agents will be good. I mean, you know, they’re agents who do autonomous work, you kind of think of them as self-contained entities, but most agents, actually, especially employee-facing agents, are really co-workers, like, they’re there to help. So it’s like having an intern sitting at the desk next to you that you work with very closely. And I think, I mean, that’s really the analogy in going forward. So someone doesn’t have a specific skill, you know, they don’t have years of experience in, you know, email management or whatever, the agent will coach them.

It will provide them the skills that they need over time. So it’s like a learning And so they’re learning on the job, but they’re being coached by… AI to be able to provide what they need to do as a human, and then the AI will do what it does. So it’s not like the AI feeling sorry for people and giving them a job, it’s actually the AI making their life easier, because they’re providing the human with the tools they need to get… to get To get this partnership working.

Russell Scherwin:

Love it. One more thread, Marty. I see a couple… after the thread, I have a couple questions from the audience here. Let’s talk… clearly, you brought up Agentforce, and clearly that’s a book you wrote, as well as a pretty important strategy based upon what I’m hearing out of Salesforce. You also said, well, you know, you could put a harness in front of any old LLM, which kind of to me sounds like, and I would agree with, the LLM becomes a high-priced commodity, and the harness becomes the doorway into enterprise context, that way applications can interact with it.

Which, A, I agree with, if you agree with, but also, more importantly, that sounds like a pretty nice strategy, or a nice position for Salesforce. And so the question for you is. in a world of Agentic, in a world of harnesses, where do you see Salesforce playing, A, and B, I’m a marketing cloud, or a sales or service cloud customer, how should I think about my relationship with Salesforce, and what I should expect from Salesforce. In this new era.

Martin Kihn:

Well, I think our goal in the beginning with Agentforce, and it’s almost 3 years old now, believe it or not, back when it was introduced, and I was at that meeting, even internally, we’re like, what is an agent? Because GPT-3 had just come out, and it was… we were thinking of LLMs in terms of chatbots, like a back-and-forth conversation. But the idea of having kind of a virtual human who’s actually doing work, and could even do work autonomously was new. And I think that’s… it’s evolved over time. The point of something like Agentforce is to make it, possible for our enterprise to build agents, to monitor, to deploy them in a way that’s trusted, that’s grounded and trusted.

And that’s not trivial at all. Like, grounded means it has access to your, you know, proprietary data. In a very careful way. And then the trusted part is even more important. The trusted part is not just permissions and governance, but it’s things like data leakage and, you know, legal liability and… You as an enterprise, like, you and I, you know, we sit down as a consumer, we ask a GPT to give us a recipe or a workout routine, and it doesn’t do a great job, doesn’t even matter. It’s like, it’s not relevant. But in the enterprise context, if you’re putting content in front of, or interactions with your customers, your entire brand is on the line.

So that’s why, you know, you need something like these harnesses.

Russell Scherwin:

I love it. So, if I’m a Salesforce customer, historically, I’m in the sales world, so I’ll think about CRM, I have bought, and I’ve been a Salesforce buyer, to make sure that I can drive solid buyer-seller interactions and own a relationship outside of the context of one conversation or one opportunity. What I’m hearing from you is. that’s great, but in a new world, your system is about, you know, orchestrating human interaction, but also bringing agents in, and what I hear you saying is it’s about building, deploying agents in a manner that they’re trusted, so they can use your application and do the things that the application monitors. in a side-by-side with human manner.

Is that… accurately represent.

Martin Kihn:

Yeah, and it… the future of work, it seems, it’s already there, actually, it’s not even the future, is this hybrid… it’s a hybrid workforce. You know, most of us are using AI in our jobs, and, like, our sales team here at Salesforce relies on AI for all kinds of things, like, not just, call scheduling, but call summary, deal points, etc, etc. So we are in that phase already of, kind of, hybrid human agent co-work, but I think that the purpose of something like an agent force is to enable the agent part of it. So you have the human part, and, you know, that’s what you run, and then we enable the agent part, and the ability to work, the humans and agents to work together.

So it’s a hybrid workforce.

Russell Scherwin:

It’s, it’s crazy to think how far we’re coming and how fast it’s.

Martin Kihn:

So fast, I know. It’s crazy.

Russell Scherwin:

So, two questions to see from the audience, Marty. Here’s the first one. Do you see a system harness and then an array of domain-oriented harnesses?

Martin Kihn:

Oh, I see. That’s a good… yeah, the answer to that is yes. I think the analogy would be, you have an orc, like a… a lot of our customers now, they’re building agents, and then a group of agents that work together, and so if you do that, in order to coordinate the agents, you can quite often build an agent on top of these other agents, so you become, like, a master orchestrator agent and the sub-agents. You could do that with harnesses as well. You might have, like, a marketing-specific domain harness, I suppose, and then your master enterprise harness, and It can get very complicated.

I think there’s no limit. If you have the engineering expertise in-house, which a lot of people do, a lot of people don’t, a lot of people do, then, you know, you can construct all kinds… whatever works best. But the harness is important. I think there’s a lot more emphasis on it now. I don’t think the LLMs become commodities. They’re the essential ingredient, like the… the brain, but I think a lot of the, kind of, competitive advantage and future will come from Harness deployment.

Russell Scherwin:

Agreed, and I love the follow-up question. By the way, whoever in the audience asked this question, whatever you were doing, you’re thinking what I was thinking. The question is. Marty, can you speak to the racy for this factory? You know, does the CI own it, and the business contributes? Going back to that decentralized execution, centralized control, you know, who’s responsible, who’s accountable, who’s consulted, and who’s informed? How does this Conceptual structure get built.

Martin Kihn:

I think it… I mean, it’s turned out to be important that you do… you do have a team that is your kind of AI center of excellence. So, the reason for that is you want to democratize these tools. You want them to be in the hands of as many users as possible that make sense for your business. kind of democratize ideation and agent creation, but the AI part itself, like, which model to use in what context, cost control, which harness to deploy, all that is… it’s a technical conversation, and it requires people who know quite a lot about AI and the foundational models and so on.

So you do… you do need this kind of central… I don’t know what you call it. It could be this… it could be the… CIO. Absolutely, I’ve seen companies where that’s the case. Sometimes there’s a new role created, like chief AI Officer, but there’s always a group of people who’d be, either, whether they’re called it or not, they’re the center of excellence around AI, so that’s important. And they provide, like, the guideposts, the guardrails. They may not have absolute, absolute decision-making capabilities, but they’re essential, and then it trickles down. I think it is technical still.

AI will get less technical, probably, but it’s technical now, so a lot of the responsibility sits in IT. It’s just a natural place for it. That’s where the engineering team sits, usually. And then, somebody like marketing would be a consumer, and they can have input, of course, and marketing’s more technical than it used to be as well. So, you know, there are hybrid teams, marketing and IT, working together. But I think, the internal structure doesn’t change that much, with the exception of this new… this new group, the AI. kind of gurus.

Russell Scherwin:

I really would love that to be a follow-up conversation, because anytime a transformational technology’s been introduced, it’s the organizational governance that’s been the lever for profitable growth. But I do see 9.30 Pacific on my watch, so Julia. I am going to be a good steward of time and hand it back to you. Thank you so much. Well, and Marty, this is an awesome conversation. I… I love it.

Martin Kihn:

Yeah, great talking to you, Russell. Yes, great topic.

Julia Nimchinski:

Phenomenal conversation. Thank you, Martin. Thank you, Russell. What’s the best way for the community to support your work?

Martin Kihn:

Me, personally, follow me on LinkedIn. Actually, I just launched a series, it’s kind of an AI explainer, how do LLMs work, so I recommend that on the YouTube channel. And then you can follow Agentforce as well on LinkedIn.

Julia Nimchinski:

Awesome. Russell?

Russell Scherwin:

Yeah, for me, say hi to me on LinkedIn, follow me there. Also, actually, I’m building my site with Claude, so, it’s in design review mode right now, which is really crazy. So go check out B2BTracks.com, and if you have comments or… pick a version, tell me which one you like.