First in a series of articles and now in multiple interviews.
This is the third and final part of Bryant's exclusive video interview. farm tech® covers the processes known as Human in the Loop (HITL) and Human on the Loop (HOTL) and how they relate to AI workflows. Bryant uses an analogy to air traffic control.
“Think of the agents in a multi-agent system as airplanes,” Bryant said in an interview. “So the agent is the airplane, and the agent has the autonomy to act on its own. The orchestrator in this analogy is air traffic management. So it's the coordination layer that sets the rules, sets the constraints, shares the context, and deals with all the airplanes in the air at a volume, speed, and complexity that's beyond human capabilities. So we have complex systems that allow the airplanes in the air to do what they're supposed to do.”
Bryant explains that HITL does not replace HOTL in this analogy. HITL remains.
“There are intentional points in the process of how we design humans. For example, in this case, the pilot is a human, and the pilot flies the plane,” he says. “They take off and land the planes. They're there. That's their role. But with HOTL, you not only have the pilots who can react to the unexpected, but you also have humans in the control tower who actually control the air traffic management system itself.”
Part 3 of Bryant's interview can be viewed above. see first part
3 articles written by Bryant are available
transcript
Editor's note: This transcript is a lightly edited version of the original audio/video content. It may contain errors, informal language, or omissions spoken in the original recording.
Most of what is currently being sold as agent AI is either steam or, frankly, fake. So the change in thinking and design here is important. Must be designed for open platforms. You need to design for interoperability and priority discovery, and building a walled garden now limits those possibilities. This will limit the entire company. You're going to really shoot yourself in the foot later on.
So we start with open platform thinking, shared orchestration, shared governance, shared agent capabilities. It's all about sharing and thinking, and the platform will be designed with interoperability in mind. Therefore, we want to avoid traditional small, departmental mini-AI ecosystems. In fact, it becomes technical debt.
So internally, I think it's important to recognize that teams within these gardens build their own agents experimentally. Either it's healthy or you want to promote an app store of agents within your enterprise for internal use, and out of that emerge the best agents that solve specific problems, elegantly recurring frustrations. Again, this is useful, but real corporate value here requires connective tissue.
Data quality remains important, and one of the biggest reasons for this is not just poor inputs and outputs. But when you're talking about systems with these controllable levels of autonomy, without that data quality, you're not only adding value, but you're potentially introducing errors as well. They can be compounded.
HITL (Human in the Loop) is designed to insert a human into a process at a fixed location. This is a design choice, but HOTL (human in the loop) is dynamic, so ultimately the human is in control of the system, but it is also invoked by the system, for example, if there is a certain risk or certain uncertainty, or if an anomaly is occurring.
So I think the air traffic control model, which obviously has been around for a long time before GenAI, captures this concept very well. For example, think of the agents in a multi-agent system as airplanes. In other words, the agent is an airplane, and the agent acts with its own autonomy. The orchestrator in this analogy is air traffic management. It's a coordination layer that sets rules, sets constraints, shares context, and processes all these planes in the air at a volume and speed and complexity that is beyond human capability. That's why we have complex systems that ensure that planes in the air continue to operate as intended.
Now, the Model Context Protocol (MCP) is, frankly, how you connect the data that is being shared. So these planes share data about their position, speed, speed, identity, who they are, and this is the common layer that allows these planes and the orchestrator to understand each other. Then there's the protocol. It's a data sharing protocol. This is truly a universal adapter for data.
Rather than agents connecting to data, there is a separate protocol for how agents connect to each other. The analogy here is collision avoidance. In effect, planes can negotiate directly in the absence of humans. You go up, I go down. And it takes place within this limited autonomy, within the rules of the system.
Now, about HITL and HOTL, you asked about the version of HITL here. This is because HITL does not replace HOTL. HITL remains. There are intentional points in the process of designing humans, for example in this case the pilot is a human and the pilot flies the plane. They take off and land planes. they are there. That's their role. But in HOTL, not only do we have pilots who can react to unforeseen situations, but we also have humans in the control tower who actually control the air traffic management system itself. This means humans are monitoring the big picture and will intervene if thresholds demand it. And this is the model for agent AI in the pharmaceutical industry.
