Every enterprise talks about agents and autonomy, but very few have moved beyond copilots taped to legacy workflows. Brayden Levangie is the exception. At twenty, he is building an architecture that turns language models into self-learning digital colleagues. It is the closest thing we have seen to the HFS vision of Services-as-Software delivered for real.
In this interview, David Cushman, Executive Research Leader at HFS Research, speaks with this 20-year-old prodigy whose company, Levangie Labs, is building what Brayden calls his “cognitive architecture” – a platform delivering genuinely autonomous agents that can learn, reason, and act in the world.
In this conversation, Brayden uncovers the thinking behind a platform that replaces scripted automation with systems that grow and discover better ways to work. If you want to understand the future of autonomous enterprises, this interview is your starting point…
“I didn’t want to chat with GPTs, I wanted them to build things”
David Cushman: What’s your breakthrough idea?
Brayden Levangie: It came from years of experimentation. When I was about 13 I got into a summer program at MIT where we played with primitive language models such as GPT-2. I later managed to get private access to GPT-3 by just emailing OpenAI – back when they were small enough that someone would answer.
I didn’t want to “chat” with it, I wanted to build things. One of the first projects I published online became an early instance of what people would now call a retrieval-augmented generation (RAG) system, though no one was using the term then. I just wanted to make an AI that could answer questions factually.
At the same time, I was obsessed with robotics. I built facial-recognition engines at Lincoln Labs and tried to embody intelligence to make it experience the world. Those experiments became the seeds of what we now call the cognitive architecture; the culmination of seven years of research and building.
David Cushman: Who backed you through that journey?
Brayden Levangie: Nobody. I was self-funded. My first “VC” was mowing lawns for $100 a month and helping out at a retirement home. Later, when I was 17, a New York startup hired me as lead AI engineer after seeing my projects online.
From chat to action: breaking the conversational consensus
Brayden Levangie: Most people equate language models with chat because ChatGPT trained the world to think that way. But chat isn’t action. Systems optimized for user engagement are not optimized for work. They keep coming back to you for another round of conversation. It’s like hiring someone who never stops talking and never delivers.
We flipped that paradigm. Our cognitive architecture sits on top of existing LLMs, from Anthropic, OpenAI, and others, but changes how they behave. Instead of optimizing for dialogue, we optimize for objectives and outcomes.
When you seed an instruction, you’re not chatting with the LLM; you’re triggering what we call an autonomous reasoning loop. The system talks to itself, plans, acts, and learns until the objective is achieved.
That’s what makes it different from the “wrappers” you see everywhere. Those are just tool-calling layers glued onto chat APIs. We’re rewriting the behavior of the underlying model.
Agents learn from experience – remembering what matters, when it matters
David Cushman: Everyone claims their multi-agent system “learns from experience.” Does yours really?
Brayden Levangie: That’s a common misrepresentation in the industry. Most so-called “learning” is just RAG, remembering a few facts or preferences and replaying them later. We’ve gone beyond that with what we call an episodic memory system.
Instead of memorizing rules, our agents form experiences and learn from them like humans do. Imagine you give a presentation and someone tells you afterward you made a mistake. Next time you prepare, that feedback surfaces automatically. That’s how our agents operate. They can back-propagate through past experiences, recognize where they went wrong, and adjust future behavior.
It’s neuro-symbolic: blending deep-learning perception with symbolic reasoning. That’s why we call it the cognitive architecture. It learns through experience, not through reinforcement rewards or pre-programmed instructions.
A new form of intelligence that can operate in the digital and physical worlds
Brayden Levangie: Reinforcement learning is like training a mouse to press a button for cheese. The mouse never knows why the button matters. Most agents work that way, rewarding signals without understanding.
We removed the reward altogether. Our systems learn from outcomes and context, not from external scoring. They gain understanding from experience. That’s what lets them operate both in the digital and the physical world — from patent law to humanoid robotics — without us pre-programming every move.
Real-world disruption: from patent law to venture capital
David Cushman: Give me an example that makes this real for enterprise leaders.
Brayden Levangie: A Silicon Valley IP and patent-law firm gave one of our agents a challenge. Our agent read a book written by the firm’s founder on patent law, received minimal feedback, and then solved complex casework at a quality comparable to a partner with several years of experience. It literally taught itself how to practice patent law.
In another case, a climate-focused VC firm used our system for market analysis. After a few feedback rounds, the agent not only completed an industry report but predicted the exact company they were about to announce investment in the next day, it had become that perfectly aligned with their thesis. That was months ago; the framework is far more advanced now.
Architecting intelligence that builds itself
Brayden Levangie: The next leap is automation of the automation. We built an Agent-Creation Agent; a meta-agent that designs new agents for specific clients or domains. When it deploys into an organization, it learns on the spot from the people who work there.
That’s how our clients, from construction to robotics to enterprise software, are deploying self-evolving systems that adapt to their culture and workflows.
David Cushman: How would it, say, design a new market-growth program?
Brayden Levangie: You’d simply talk to the Agent-Creation Agent, describing your goals in free form. It builds a new intelligence inspired by your intent. Because it learns from your thought process, it can even come back with better strategies than you initially proposed. Many of our breakthroughs emerged that way, when the agents themselves go beyond the brief. When you give something the ability to learn, you also give it the ability to discover better.
Working with (not for) the big bucks LLMs
David Cushman: So where do OpenAI or Anthropic, for example, fit into this picture?
Brayden Levangie: We do call their APIs, but only for a small part of the process. The heavy lifting; reasoning, memory, learning, all that happens inside our architecture.
Think of it as using an LLM’s ability to generate possible next tokens as the raw material. We harness that, route it through our cognitive and memory layers, and the agent decides what to do next.
We can even run it on-prem with licensed LLM weights when privacy is critical. Some partners, including big tech names you may be familiar with, are already doing this with us under NDA.
The result: a lower-cost, higher-value system that delivers outcomes rather than conversations. One of our first commercial applications was the world’s first autonomous patent-agent. It performs end-to-end patent filing with no human in the loop beyond initial guidance.
True autonomy — with humans as creative directors
David Cushman: This sounds like what we at HFS call Services as Software: humans defining outcomes, software delivering them. How far are we from that?
Brayden Levangie: We’re already there. Our agents operate genuinely autonomously with no new human input once you set the goal. But the human still defines the goal. That’s why I use the term Creative Director.
Humans provide vision, intent, and passion; the “why.” Agents handle the “how.” In my own company, agents handle much of the engineering, marketing, and business ops, allowing me to focus on strategic direction and partnerships. My job is to be the creative director, setting direction and ensuring alignment. We’ve effectively become one of the first autonomous organizations.
Now you can build systems that automate discovery itself
David Cushman: How do you plan to monetize this?
Brayden Levangie: Carefully. It’s too powerful to just throw into the wild. Right now we work with select high-impact deep-science firms, advanced-tech startups, and IPO-bound companies, that can use it responsibly and at scale.
Our vision is not to be another B2B SaaS agent platform. We’re building a system that automates the process of scientific, technological, and creative discovery. Humanity needs acceleration in all of those areas to solve its biggest problems. These agents can help us do that.
Yes, it’s a for-profit company, but profit fuels progress. We’re aligning with partners who share a public-good mindset. In the long run, this becomes an infrastructure for collective progress, not just another enterprise app.
Replace sunk-cost failed AI with full autonomy
David Cushman: What kind of companies make the cut?
Brayden Levangie: We’re not short of interest, so we’re picky. The number-one criterion is alignment. I have to feel I want to work with the founders. Culture matters even in automation.
Mostly we’re partnering with technology-centric enterprises spending millions on AI projects that our agents can replace or outperform quickly. They come to us saying, “We’ve sunk huge budgets into AI that still needs humans in the loop.” We show them what full autonomy looks like.
David Cushman: Enterprises still have to buy foundation models from the big players, right?
Brayden Levangie: Sure. But we’re not competing with the LLM providers; we’re complementary. They supply the raw linguistic intelligence; we supply cognition, memory, and autonomy. Think of us as infrastructure-layer innovation, not application-layer AI. We’re re-engineering behavior at the token-generation level, turning probabilistic text prediction into purposeful reasoning. That’s what turns language models into agents that act.
The next technological epoch offers systems that grow
Brayden Levangie: Every day the architecture improves itself. It learns new domains, designs new agents, and contributes back to our internal ecosystem. We’re watching intelligence compound in real time.
For the first time, we have a system that can grow not just run the code we wrote yesterday, but write better code tomorrow. This is the next stage of technological evolution.
Humanity has always accelerated progress by creating tools that amplify labor. Now we’re creating entities that amplify thought.
Humans remain firmly at the helm role in the autonomous age
David Cushman: And what about people’s jobs? This sounds like a lot of humans out of a lot of loops?
Brayden Levangie: The creative-director model keeps humans essential. The systems execute, but humans define value, ethics, and purpose.
In my view, we’re moving from labor-based organizations to imagination-based ones. The winners will be those that learn to orchestrate fleets of autonomous agents toward bold human goals.
DC: Brayden, you’re 20. You’ve been building this since 13. Do you ever step back and think: this is moving fast?
Brayden Levangie: Every day. I built most of this in a spare room in the woods of Massachusetts. Now I’m in San Francisco, ready to shape a generational shift, building the future on my own terms, and hopefully for the better of everyone.
Bottom line: The services-as-software inflection point is here. Autonomy means services can be fully delivered by software.
Where many of today’s AI “agents” are scripted copilots, a new era of self-evolving digital colleagues takes us on a leap from automation to autonomy – delivering the inflection point at which services can be fully delivered by software. Prepare to redesign your organisations with humans as creative directors guiding fleets of intelligent agents toward business transformation – with the powerful benefit of the daily, autonomous, discovery of better.
Brayden Levangie’s cognitive architecture is at the leading edge of the shift to full autonomy. Cognitive architectures, episodic memory, persistent state, and autonomous loops, are now in the mainstream of cognitive-architecture development and agentic-LLM thinking. The result is a leap forward in working, multi-domain, persistent agent systems that enterprises can use in anger.
Demos that Brayden has shown HFS suggest an integration level and autonomy to compete with the most advanced commercial agents – such as Devin in the world of coding agents. Levangie Labs application in construction, spatial reasoning, legal IP, and investment also indicate the framework is broadly applicable across verticals and enterprise use cases.
Posted in : Agentic AI, Artificial Intelligence, GenAI, Generative Enterprise, Services-as-Software
