Reagent Systems LLC | AI Research, Agentic Apps

What is the AI boundary?

The AI boundary is the gap between what we want computers to do for us and what computers are currently able to do. It's been a little over 40 years since the first easy-to-use computer was introduced to the public, and since then software has been torn between ever-complicated systems and easy-to-use interfaces. As systems become more complicated, it makes sense for these systems to be maintained and managed and interacted with, through another system. That's where an AI Agent comes in.

What is an AI Agent?

Think of an agent as an intelligent computer wrapped up as a software package. Instead of you dealing with complicated software, the software deals with itself for you. A fully autonomous AI Agent is intelligent enough and capable enough to design any workflow it needs to accomplish a task. This is a new paradigm for interacting with software, but you can simply think of it as your computer doing any task you give it!

How close are we to fully autonomous AI Agents?

Pretty close! Members of our team led the research group to create the world's first AI Agent, and we pushed the boundary further than ever before. Multi-agent systems, emergent engineering, evolutionary algorithms- you name it, we've been using it. We're still pushing the boundary today, with all of our existing apps and more coming out for you to try!

Will an AI Agent do what I want it to do?

Our goal is to make AI Agents do everything you want it to do. If you find it's not up for the task, please let us know, as our goal is push the boundary, and you found one!

Team

Meet the people behind Reagent Systems.

Kyle Steel

@ThyFriendlyFox

Chief Executive Officer 27 • Florida

I graduated from Florida Polytechnic University with a Masters of Engineering in 2022 with a thesis focused on using Fourier-based machine learning techniques to calculate nonlinear physics systems with greater precision and speed on an Arduino running a custom algorithm than the available university supercomputer running Ansys Mechanical. My passion has always been in Artificial Intelligence, and I even led the research and development team for AutoGPT for a short period of time, unlocking the possibility of truly multi-agent systems. Since graduating, I've been an ITAR and CMMC compliance system administrator. Most of my work has been developing a CMMC compliant ERP system for the manufacturing industry. I'm a self-taught software developer and system admin with experience running businesses for a number of years now.

Ethan Shelton

@Spike

Chief Information Officer 25 • Florida

Hello, I'm Spike! I've been coding since I was 13, diving into mods, mobile games, VR experiences, websites, AWS, and software development. I thrive on the challenge, the puzzles, and the fun of piecing it all together. Currently, I'm focusing on my vlogging and VR YouTube channels and crafting my own AI assistant named KIT. This is only the beginning to my story of relentless creation.

BentlyBro

Chief Technical Officer

Hey, I'm Bently — a self-taught developer, systems tinkerer, and all-around builder with a love for pushing boundaries. Whether it's designing encrypted chat platforms, wiring up brain-computer interfaces, or spinning up AI agents that talk to each other, I'm always crafting something weird, useful, or both. My portfolio (which totally needs an update) showcases a mix of tools like Promptly—a live website builder powered by natural language—and NoteNavigator, an AI music recommender that ran a little too close to Spotify's sun. As the community lead and developer behind AutoGPT, I balance hands-on coding with leading a vibrant Discord full of chaotic brilliance. Most days you'll find me deep in Python, JavaScript, or building cross-device systems that just sync. Other days, I'm designing VR worlds, experimenting with EEG data, or mapping out new ways to bring decentralized, real-time tech to life. Welcome to my world—it's a bit experimental, often chaotic, but always driven by curiosity.

Colton Frear

Chief Financial Officer Florida

My name is Colton Frear, I graduated from Florida Polytechnic University with a Masters in Mechanical Engineering. My first job out of college was as a Liaison Engineer for General Dynamics Electric Boat In Newport News Virginia. While at Electric Boat I provided engineering support on the Virginia Class Nuclear Submarine program for NAVSEA. After a short time there, I moved to Palmdale California to work for Northrop Grumman Aeronautics Systems. During my time with these prime contractors I have become intimately familiar with DoD program requirements and expectations as well as improving my general technical and engineering knowledge.

Alexey Kuznetsov

Chief Science Officer 25 • Lakeland, Florida

I graduated from Florida Polytechnic University with a Bachelor's in Computer Science focused on cybersecurity in 2022, followed by my Master's in 2024, with plans to complete a Master's in Data Science by 2028. My biggest achievement has been co-developing Dynamo, a comprehensive all-in-one system built from scratch to run machine shops of any size, using JavaScript, HTML, Python, and PowerShell scripts. What started as a challenge to modernize manufacturing operations has grown into a full-scale platform that I continue to maintain today. My mix of formal education and self-taught skills has shaped me into a versatile problem-solver who thrives on building practical solutions. I bring my cybersecurity background and hands-on development experience to tackle whatever unconventional challenges come my way. Whether it's maintaining enterprise systems or exploring new technologies, I'm driven by the satisfaction of creating tools that actually make a difference.

vision

We envision a future where AI systems work seamlessly with humans to enhance our capabilities and solve the world's most pressing challenges.

research

Decentralized Federated Learning Network

We propose a decentralized federated learning network to train a 1.58-bit FP4 quantized language model based on Hinton's forward-forward algorithm, trained on function calls. This approach enables training language models that are far cheaper, faster, and more private than traditional methods.

Key Innovations

Forward-Forward Algorithm: Replaces backpropagation with dual forward passes, enabling greater parallelization of Large Language Model training
Ternary Weights & FP4 Quantization: Extreme compression (log2(3) ~1.58 bits) achieving 10-100x cost reductions while maintaining accuracy
Federated Learning: Decentralized training on edge devices, keeping data local and preserving privacy
Function Call Training: Specialized datasets generated from application usage patterns

Advantages Over Traditional Methods

Unlike traditional transformer-based LLMs that rely on backpropagation, high-bit precision (FP16/32), and massive centralized GPU clusters, our approach eliminates backpropagation entirely, uses extreme quantization, and decentralizes training via federated protocols. This results in:

10-100x cost reductions compared to centralized training
Faster convergence through parallelism
Inherent privacy by keeping data local—no raw data sharing
Ability to run on edge devices, mobile processors, and embedded systems

Potential Applications

Our model architecture can be embedded into systems ranging from microcontrollers to cloud-based applications. By making the model incredibly small and accurate, we can extend AI capabilities to smaller devices where competition remains low due to limited compute. Mobile devices are a primary target, where our model runs faster, maintains privacy, and is dedicated to function calls—essential for effective AI agents.

Status

This research is currently under review through the National Science Foundation (NSF) SBIR program. Our team, originating from the AutoGPT research and development lab in 2023, has developed three applications, three agents, and a framework for training LLM models on Apple Silicon hardware.

Our research focuses on advancing the frontiers of artificial intelligence, machine learning, and human-AI collaboration.

Our current research is undergoing review processes. If you'd like to learn more, please get in touch with us.

products

From idea to implementation: Our product creation process and top solutions

Our Process

Discord Hub

Our Discord server serves as the central nervous system of our community. It's where real-time conversations happen, where quick questions get immediate answers, and where spontaneous collaborations are born. Members share breakthrough moments, seek immediate feedback on prototypes, and coordinate collaborative efforts.

Creative Forum Board

This is where imagination runs wild. The Creative Forum Board is designed for exploration, brainstorming, and early-stage concept development. Here, members share raw ideas, experimental thoughts, and "what if" scenarios without the pressure of immediate implementation.

Project Forum Board

When ideas are ready to become reality, they graduate to the Project Forum Board. This space focuses on planning, execution, and refinement. Here, concepts transform into actionable roadmaps, prototypes, and eventually polished implementations.

Validation & Assistance

Our forum systems provide immediate feedback loops that help members understand the potential and practical challenges of their concepts. Our community includes experienced developers, researchers, and entrepreneurs who can quickly identify both the promise and pitfalls of new concepts.

Implementation

From concept to implementation, our community infrastructure supports the complete lifecycle of innovation. Whether you're a seasoned professional or someone completely new to the field, our systems provide the support structure needed to transform concepts into reality.

volunteer

Join our community of researchers, developers, and AI enthusiasts working together to advance the field.

donate

Support our research and development efforts to advance AI technology for the benefit of all.

What is the AI boundary?

What is an AI Agent?

How close are we to fully autonomous AI Agents?

Will an AI Agent do what I want it to do?