Autonomous Agents 101: How Industrial Engineers Can Teach Machines to Solve High Value Problems
Above: Hunter Park (left) and Octavio Santiago.
A Deep Dive with Octavio Santiago, Head of Data Science, and Hunter Park, Lead Machine Learning Engineer at Composabl
Foundational industries like manufacturing, energy, and supply chain are the backbone of the global economy. Yet, they face a growing challenge: how to become more efficient, resilient, and sustainable in a world of increasing complexity. From optimizing production lines with thousands of moving parts to managing intricate energy grids or supply chains spanning continents, the sheer scale and speed of these operations demand smarter tools to assist human decision-making.
This is where autonomous agents and multi-agent systems come in. These artificial intelligence (AI)-powered systems are designed to augment human capabilities—analyzing vast amounts of data, making informed recommendations, and taking actions in real-time to optimize processes. Imagine a factory where engineers rely on machines to detect potential equipment failures early and adjust operations seamlessly to maintain productivity. Or a supply chain where logistics teams work hand-in-hand with AI agents that dynamically reroute shipments during a storm to prevent delays.
Autonomous agents don’t replace human ingenuity—they amplify it, enabling people to tackle the kinds of challenges that once seemed insurmountable. By bridging complexity with clarity, these tools empower industries to thrive in an increasingly demanding world.
At the forefront of this revolution is Composabl, a groundbreaking platform that enables engineers to design and deploy multi-agent intelligent systems. Backed early by Ridgeline, Composabl exemplifies the kind of innovation needed to modernize foundational industries and secure their future. What sets Composabl apart is its focus on multi-agent autonomous AI systems—systems that not only make decisions but continuously learn and adapt to optimize complex industrial processes in real time.
To delve deeper into this transformative technology, we spoke with two of Composabl’s key leaders: Octavio Santiago, Head of Data Science, and Hunter Park, Lead Machine Learning Engineer. They shared their insights into how autonomous agents have evolved, the challenges of building tools for such complex systems, and the extraordinary real-world applications already making an impact.
In this Q&A, we explore Octavio and Hunter’s perspectives on the future of AI and automation, and how Composabl is helping reshape the industries that keep our world moving.
What are multi-agent autonomous AI systems and how would you explain that to someone new to this space?
OS: The concept of multi-agent autonomous AI systems is simple yet powerful: it’s an AI system designed to act, not just think. These multi-agent systems use a methodology called machine teaching and a combination of algorithms to solve highly complex processes with thousands of sensor variables and actions, as we usually have in real-world scenarios. The objective of these agents is to find the best actions that will improve the business processes through data and control the system with these actions in real-time.
HP: AI agents have received a lot of buzz around the industry in the last year or so. For Composabl, agents go much deeper than taking an off-the-shelf LLM and using pre-existing tools to loosely fine-tune it to poorly fit whatever use case companies force into products. We believe that multi-agent systems are required to truly drive the solutions manufacturing and large industrial companies need, there is no one technology or genAI tool that solves it all, so our customers need the ability to orchestrate multiple technologies and agents to get the autonomous decision-making results they need.
Autonomous agents are fundamentally grounded in statistics and features. In order for autonomous agents to truly thrive, there has to be human understanding as to why they are making decisions: interpretability. This interpretability cannot come from the same source as the agent. In other words, it is a logical fallacy to both ask an LLM agent to do a task, then have it explain its reasoning. It is circular reasoning, and ultimately the LLM evaluation then includes how well it can explain itself, and not rigorously grounded in empirical results verified by tried-and-true data science practices.
Multi-agent systems are made to overcome variety in the environment they operate in. They need to be deterministic and auditable.
For people who are just learning about this technology, the idea of machines making decisions autonomously might feel like a giant leap. How would you explain to them the progression from early AI systems to the autonomous multi-agent systems we see today?
HP: The Composabl Platform’s purpose is to make sequential decisions that have cascading, often nonlinear, effects to the larger system it is operating in thereof. Early AI systems, pre-2012 in particular, had major difficulties learning these hidden states of systems without having them be explicitly modeled by an engineer of the system. Without a true understanding of the causality of actions, there is no hope for any autonomous system to operate with the complexity and scale we see Composabl agents deployed in today.
Modern autonomous agents are experts at cause and effect, and Composabl’s autonomous agents, orchestrated within the multi-agent system, provide the ability to learn and operate within these dynamic systems and are more rigorously defined and optimized than any other agentic solutions.
OS: Real-life decision making problems are complex and require significant computing power and data to solve them in real time. While early AI models are focused on batch predictions without direct interference in the environment, autonomous AI agent models work with real-time predictions and decision making models. Autonomous AI agents used to be a combination of algorithms and solutions into one agent that is deployed in a real process to optimize the decision making by taking actions into the system. The increase in computing power in recent years and the advancement of automation have made this transition possible. Today, companies can use hundreds or thousands of sensors to train robust AI models and enable the deployment of multi-agent AI systems in real-time operations.
The pace of innovation in AI has been extraordinary in recent years. From your vantage point as leaders in machine learning and data science, what are the biggest breakthroughs that have made today’s capabilities possible?
HP: The recent years of AI can only be described as the Second Great Awakening (the first being 2012-2015 with deep learning). The most important innovations are not tokenization improvements for Large Language Models (LLMs) or Retrieval-Augmented Generation (RAG), but rather the continued mathematical understanding into neural networks as a function of inputs and outputs. The most important papers rigorously prove how automated feature learning comes to be, and why the features learned are so expressive and span-filling.
Composabl as a Platform stands on the concrete edifice composed from these papers and over a decade of empirical results, whereas pure LLM approaches are a generator of probabilities not provable in anything.
OS: The biggest breakthroughs were the advancement of data-driven culture in the industry, which enabled the execution of more innovative projects, investment in high-performance computing, and the structuring of specialized teams. This culture was driven and paved by important concepts such as data science, cloud computing, Industry 4.0, digital twins, online learning, among others that emerge every day. With all this effort in AI development today we can develop and implement advanced AI solutions such as autonomous AI agents more easily and efficiently every day.
Composabl’s platform empowers engineers to solve nuanced, real-world industrial problems. What unique challenges have you encountered in creating tools that simplify such complex processes, and how have you overcome them?
HP: My favorite tool that empowers engineers to solve problems, whose prospect of automating was a distant dream only a few years ago, is the Perceptor. As someone who has developed many Machine Learning (ML) solutions, both professionally and as a hobby, the most frustrating part is trying to train a model that is struggling to learn a feature that any human would be able to identify immediately. Instead of blindly relying on whatever exploration employed to stumble upon scenarios where that feature would be useful and further hoping that the model will eventually learn them, the Perceptor explicitly allows the engineer to provide these latent features and greatly expedite training.
Only recently has this exact technique been published by The AI Institute, the AI research arm of Boston Dynamics. It was heralded as the next greatest thing, but Composabl’s perceptors (for example: vision, seismic, and temperature gauges) exist in the real world, not just in labs.
What are some of the most exciting or impactful real-world use cases you’ve seen with autonomous AI agents, and how do they illustrate the potential of this technology?
HP: Here are some incredible use cases:
A CPG beverage producer deployed an agent to control a line of machines producing aluminum cans. Their goal was to improve plant throughput by 1%, and in live factory tests, the agent ended up delivering a 3% improvement.
A chemical company developed an agent to calibrate a thermal reactor for a new product introduction. The agent reduced the calibration time from six months to just 72 hours.
In healthcare, autonomous agents have been used to improve diagnostics and reduce readmission risks. This is the most exciting real-world use case for me personally. These systems uncover hidden relationships in data and perform complex statistical analyses with far greater precision and speed than manual methods. By identifying novel correlations, they’ve significantly improved patient outcomes, demonstrating the transformative potential of AI in medicine.
Can you share an example of how Composabl’s platform has been used to solve a complex, real-world industrial problem? What lessons did you learn from that experience?
HP: One of the most compelling examples of Composabl’s platform in action is with a leading glass bottle manufacturer. They developed a multi-skill AI agent using our platform, designed to tackle various tasks seamlessly by employing a selector design pattern. Each skill within the agent is highly compact—just 500MB—making it incredibly lightweight and efficient.
What’s remarkable is that this trained agent can run effortlessly on a device as small and low-powered as a 5V Raspberry Pi. This stands in stark contrast to large language models (LLMs), which require multiple GPUs and consume hundreds of watts of power. The lesson here is clear: advanced AI solutions don’t need to be resource-intensive to deliver impactful results. By prioritizing efficiency in both model inference and feature compression, we’ve enabled a practical, scalable solution that’s accessible for real-world industrial applications.
OS: In the impactful use case Hunter described above, the AI Agent designed by the company’s engineers masterfully controlled an extremely complex industrial process, where the goal was to increase production while complying with product quality parameters. This industrial process consisted of a system with more than 90 sensors, and the agent made decisions through more than 50 actions in the process. In addition, the model was trained on a data-driven digital twin, made 100% through historical process data and machine learning models.
Automation has become a cornerstone for reimagining foundational industries like manufacturing, energy, and supply chains. Why do you think automation is critical for their future, and how does Composabl fit into this shift?
HP: Automation is the bedrock of manufacturing, energy and supply chain industries. The cost of automation is great, the cost of not automating is greater. Once something is automated, the robots built to particular form factors, the control programs coded, it is very hard to make changes that will yield improvements to the bottom line. Composabl is precisely the platform to discover these optimizations to pre-existing automation solutions to maximize the efficacy of the system.
Hunter, what drew you to this field, and how has your approach to problem-solving evolved as you’ve stepped into this role at Composabl?
HP: My attraction to the field of machine learning, as a whole, started with traditional computer vision research for mobile robotics. After doing hand designed feature engineering so many times, you begin to wonder if there is a way to automate it. Right around this time, AlexNet came out and showed the world what Convolutional Neural Networks could really do.
During my tenure at Composabl as Founding Machine Learning Engineer, it always offers an exciting challenge. The problem solving is not on a problem basis, but on a problem-type basis. The Composabl platform needs to be coded in such a way that it not only is capable of solving one problem, but entire classes of problems without issue. This requires thinking on another level of abstraction that is both challenging and very rewarding as new customer projects get introduced and we are confident in the Platform’s ability to solve it already.
Ridgeline’s thesis emphasizes updating foundational industries through automation and AI. From your perspective, what do these industries need to understand about staying competitive in a rapidly evolving technological landscape?
HP: Some organizations might feel nervous about the utilization of AI with their pre-existing systems. It is not uncommon for industrials to have low levels of (and sometimes near zero) infrastructure that easily enables an AI system to observe and fine-tune that organization's processes. However, the good news is that the barrier to entry is diminishing rapidly, and companies can drive their competitive edge without driving or blowing up their costs.
If a company wants to utilize AI, the great news is that with the Composabl platform they no longer need to hire a team of machine learning practitioners or data scientists. They mainly need to focus their time and energy on being able to iterate quickly and ultimately empower the operator and process engineers, who built the physical systems, to distill their deep knowledge of the problem into the AI training process.
OS: In an increasingly competitive world, having 1% improvement in a process can be the key to a company's success or failure. Technology has advanced very quickly and companies are increasingly adopting a data-driven culture that is open to innovation. A company that misses this movement may not have the opportunity to reverse it in the coming years. Automation and AI is the path to a new era of prosperity that will take companies to a new level.
““I challenge anyone to bring us an industrial problem that our system cannot solve.””
The vision of empowering 100 million industrial engineers globally is bold and inspiring. What steps are you taking to make this vision a reality, and how do you measure progress toward that goal?
HP: The prospect of empowering 100 million [industrial] engineers excites me to no end. The most crucial step of seeing this become reality is getting them to view problems not as systems of equations, but as processes with defined steps and states. Once a problem is sufficiently broken down into well defined pieces, each with their own unique responsibilities, engineers are then able to come up with an application of machine teaching that works best for this particular sub problem. Modularization of processes is vital for empowering.
Finally, what excites you most about the future of AI and automation, both at Composabl and within the broader landscape of innovation?
HP: Composabl [serves] customers all doing vastly different and equally exciting and complex problems, and it couldn’t be a more exciting time to be the machine learning engineer behind the platform. The future of AI is bright, and with respect to the innovations past and present, the equations and empirical results are only going to become more inspiring and beautiful. Composabl is an engine that makes world class autonomous agents. I challenge anyone to bring us an industrial problem that our system cannot solve.
OS: What excites me the most is that AI and automation have been around for decades now and we still have a lot of room for improvement and optimization, especially in manufacturing. From the Industrial Revolution to Industry 4.0 and now the AI era, we are seeing better sensors, better data acquisition, time series databases, IoT, Data Science and ML becoming pervasive and this is just the beginning. Composabl’s platform is being built by engineers who aim to further elevate this adoption of AI and drive innovation in the industry.
Learn more about Composabl at composabl.com.