Why AI Agents Fail in Enterprise Workflows | Colt McNealy, LittleHorse | TFiR

Swapnil Bhartiya: When it comes to AI, the general perception is that building an agent is the hard part. It’s actually not. The real challenge starts when that agent is up and running, but it has to plug into your business, your apps, your people, and actually get work done. That is an orchestration problem. And enterprises have wrestled with this problem ever before AI agents even started to enter the picture. Now there are a lot of companies who are trying to solve this problem, and one of those companies is Little Horse, that has built a saddle command center to close that exact gap. And they give agents an action layer which is powered by what they call business as code. And joining me today is Colt McNeely. And joining me today is Colt McNeely, founder and CEO of Lighthorse. Colt, it’s great to have you on the show.

Colt McNealy: Thanks for having me. Sapnil, great to meet you.

Swapnil Bhartiya: For those who may not know, talk a bit about Lighthorse. Also, interesting name. Just walk us through what problem you saw in the space that led to the creation of this company and of course, share the story of the name as well.

Colt McNealy: The problem I saw when I decided to start Littlehorse is that business processes are fragmented across dozens of different systems. And those systems could be external SaaS, it could be internal microservices, it can be glue code, can be integration platforms, or it can be tribal knowledge and processes that are driven by emails and spreadsheets. So that process fragmentation that I experienced in my first job working in the real estate sector as a software engineer made it really difficult to automate business processes. And Fast forward to 2026, with the advent of AI agents, this problem has gotten far, far, far more important. Number one, because this framework fragmentation hinders the adoption of agents into mission critical workflows. And number two, because AI has made it faster to create code and enterprises are being hit with an avalanche of new systems that they have created internally and that they must integrate to externally, which has made the problem of orchestrating workflows even more and more difficult.

Swapnil Bhartiya: Excellent. Thank you. And the story behind the name.

Colt McNealy: The story behind the name of Little Horse is. Well, the official story is that it’s fast and agile. And the inside story is that I was playing a pond hockey tournament and our team was jokingly named it the California Circus. So we didn’t put our real names on the back of the jersey, we put circus animals that were closest to us. And my name means Little Horse, so my jersey said Little Horse. And I got it the day that I created the llc and I didn’t have a better idea for a name. So that’s where it came from.

Swapnil Bhartiya: And that’s how most of the companies get started, right? Whether Yahoo or Google, you know, because name is not that important. The concept, the idea, the vision is more important. Now when it talks. When it comes to vision and idea, you kind of believe that, you know, as I was stating earlier, that this is more of more or less like an orchestration problem, as you also stated, and which is not at all related to agents at all. What is the source behind that belief that it’s an orchestration problem at enterprises?

Colt McNealy: That’s a very good question. And the reason is that agents are very, very powerful on their own already. But just chatgpt.com can’t really orchestrate your business processes. You have to get it connected to the right systems at the right place at the right time. And you have to invoke that agent as part of a larger business process. So agents are very good at accomplishing tasks that take people a very long time to do. However, a task is not a business workflow. A task is part of a business workflow. And agents are most successful and most transformative when they are headless agents that are invoked by backend systems that were written by IT or engineering teams. And that allows agents to have the biggest impact on an enterprise as opposed to simply just being a personal assistant.

Swapnil Bhartiya: As AI agents move out of experimentation and into real business processes, what is the biggest challenge you are seeing? Why isn’t just building an AI agent enough to make it useful inside an enterprise?

Colt McNealy: So there’s many things that enterprises care about that agents don’t give you out of the box. The biggest one is governance. When you just rely on the agent as the orchestrator, you don’t have any guardrails for what the agent will or won’t do, other than just a bunch of hopes and skill files and props. And then you prompt your agent and hope that he’ll adhere to the guidelines that you gave it. That’s one problem. The second problem is observability. Where agents are sometimes very opaque, it’s very difficult to trace all the tools that they call and their processes aren’t actually very observable. And the third and probably most important challenge with deploying agents into mission critical processes is durability. An agent doesn’t necessarily guarantee that a process is going to execute correctly over multiple days because business processes can take a long time, some of them take months, and they can wait on external systems. Agents aren’t really built for that. Agents are built for solving tasks that require thinking. And there’s many workflows that require months that involve different agents. The difficult part is orchestrating workflows across those agents and observing what the agents did governing them. And that’s what business code does. Business as code allows you to codify a business process at a high level in natural code and then executed across various different systems. Some of those systems might be external SaaS systems, some of them might be microservices, and others might be agents. Every step of a process orchestrated by business as code can be governed, audited, traced and durably executed. So if there’s a failure, the process can retry until success.

Swapnil Bhartiya: Why do you call it business as code not farm or barn’s code?

Colt McNealy: I do live on a ranch, but we do sell Saddle Command Center, which makes it easier to ride Little Horse, which is our open source platform. But business as code is exactly what we do. And we love the term business as code because it came to us from our first customer. And it also is analogous to infrastructure as code. Infrastructure as code gives you governance, observability, a single point of truth and automation for what systems you have deployed. And business as code does the same thing, except for your business processes. It gives you an artifact that can be deployed, version controlled, tested, to give you determinism for your business workflows.

Swapnil Bhartiya: Now you call Saddle Command Center an action layer and you use the term business as code. What does that actually mean and how does it help connect AI agents to the applications, microservices, events and people that run that business?

Colt McNealy: Business as code provides the action layer to orchestrate agents. There’s the model layer which gives you the intelligence. Then the context layer gives the agents the ability to gather all of the information that they need. And the leaders in this category are Snowflake and Databricks. But the action layer is wide open. And that’s what business as code gives you. That’s what Little Horse is the leader in. The action layer invokes the agents at the right time and allows them to do less. There’s only certain parts of a business process that need an agent. A lot of things can be done with just normal integration, plain old spring boot applications and what some enterprises are running into. One mistake that they’re making that we’ve helped them overcome is relying on the agent to orchestrate the process. When you do this, you end up spending tokens, and to your point about the cost of tokens, you end up spending tokens over and over again to figure out the same steps of integration and orchestration. Over and over again, sometimes at very low rates these days, but sometimes the model will skip a very important step which is not good in compliant or overly regulated industries. In contrast with business as code, the role of the agent is constrained to only the fuzzy logic parts, the parts where you need to make a decision that previously used to take human cognition and then the parts that are deterministic and that are specified by regulations or specified by the core secret sauce of your business process. Those parts are codified in deterministic code that is version controlled and tested and orchestrated around the agent. So business as code allows you to orchestrate those multi step processes and as we see with agents, this orchestration becomes much more important.

Swapnil Bhartiya: With version 1.3 you are rolling out a lot at once. Agent creation, pre-built task workers, skills, JavaScript SDK plus a free serverless option. What is the idea behind all of this and what does it mean when we talk about developer experience?

Colt McNealy: The 1.3 release had two main goals. The first is to increase the reach to developers with our open source platform. And we did this with two main things, the first being the JavaScript SDK and the second being the free trial of the serverless cloud. We’ve done other things such as improving the data contracts for our Structs object and type schema registry, but the biggest for the open source is JavaScript SDK and free serverless offering. Now for enterprises what we want to do is create a bunch of pre-built connectors, integrations and layers above what we offer in the open source to speed the adoption of business as code at the enterprise. So what we’re really excited about with this is a no code deployable AI agent that implements the decision worker pattern I was just talking about where you constrain the agent to making fuzzy logic decisions and then orchestrate it as part of a larger workflow. That’s a pattern that we learned from our customers and it’s now a native part of our platform. Very excited about that.

Swapnil Bhartiya: Looking ahead, how do you see orchestration changing as AI agents get more deeply embedded into enterprise applications? And what role do you see for Little Horse? And if you can talk about whether there are other players also in this space? If yes, what edge do you have over them?

Colt McNealy: I see. That’s a very good question and I’ll start with our assumptions about the future. We have two main assumptions. The first is that with AI, consumers, whether consumers and people working inside an enterprise, for example, if you’re a garbage truck company, the drivers will expect and customers will expect a more and more automated and personalized intelligent experience. And this is going to put pressure on businesses to automate processes which are currently manual to create business insights from the data that is flowing through their systems. And these pressures, point number two, are going to cause a lot of businesses that previously procured their entire IT stack and wrote only the minimum code necessary to integrate these systems. They are going to start taking the plunge into writing their own software. And that is going to be accelerated by two main things. Number one is that AI makes it easier to write code. So a team of five or ten software engineers can do what used to take fifty or one hundred in the past and that switches the build-buy calculus for a lot of systems. And then second is that orchestrating these new business processes is going to be very difficult because they’re brownfield. All of these companies, these enterprises have dozens or potentially even hundreds of SaaS systems and their business process logic is scattered across all of those systems. Now, Little Horse is unique. We are the only company providing business as code solutions that allows you to work in natural code, which can be written by Claude and understood by humans through the visual dashboard that visualizes the process. Business as code. We’re the only company that does that and we can orchestrate above the mess of brownfield systems that you’ve already got. And we do this in an open source, event driven way. It’s horizontally scalable, can run wherever your infrastructure runs, and that gives us a clear advantage over the market. It’s a generational leap over iPaaS systems like Mulesoft. There are things that our customers do at Littlehorse that you just cannot do with Mulesoft, such as building your own fleet management system that’s highly integrated with your own business process. Because in the past, enterprises used to change the way their business practice works to fit the opinions of the software that they bought. Now business as code flips that on its head and allows the software to change to exactly what the business needs. And it allows software engineers to speak the language of the business so that what gets built is purpose built for exactly what the enterprise needs.

Swapnil Bhartiya: Now let’s talk about competitors, your peers.

Colt McNealy: One competitor that we run into quite often is on one end of the spectrum is Mulesoft and the advantage we have over them and other integration platforms, and BOOMI is another one, is that Little Horse just allows you to do far more. Mulesoft exists generally to connect systems and to get data from one system to show up in another. Littlehorse solves the problem of I need to do stuff that you just can’t do with SAP, because SAP has agents now that can automate things within the four walls of SAP, but they don’t really connect to NetSuite and to my own fleet management system or my own internal microservice architecture that I’ve built. Littlehorse and business as code allow you to create coherent workflows across all of that, such that the users don’t even know that there’s multiple systems behind the portal they’re viewing. And that’s something that is very difficult to achieve with systems like Mulesoft.

Swapnil Bhartiya: Is it possible for you, you may or may not share the name, to share a real world example of a company or customer using Saddle Command Center to orchestrate AI agents today?

Colt McNealy: Yes, we can. So there is a logistics company in the United States that has 100,000 pieces of field equipment and they had previously procured their entire IT stack and they had been stitching together systems like Mobi, UKG, NetSuite, Workday and others with a lightweight integration platform. But they had been hitting a wall and as they were growing their business, they had to go through growing pains with the technology that they had. And they had been changing their business practices to work with the limited software that they had. Now two years ago they decided to flip this on its head and start building their own systems. And they discovered that orchestrating workflows across this brownfield environment was very difficult. So they started working with Little Horse with one workflow and today they’ve got over 140 workflows in production, have been able to retire certain SaaS modules and have a new platform for managing their trucks and job bidding that is highly customized, built on top of business as code and allows them to move a lot faster than their competitors because it’s something that they have and no one else has in the market. Rather than waiting for a new feature that comes out of SAP and now suddenly the entire market has it, so it’s no longer a differentiator.

Swapnil Bhartiya: Colt, thank you so much for joining me and breaking down how Little Horse is rethinking orchestration for the AI agents era. Thank you for your time today and of course those who are watching, please go and check out Little Horse to see what problem they are solving. Go to Littlehorse.io and once again, Colt, thank you for joining me and I would love to have you back on the show. Thank you.

Colt McNealy: Thank you so much.