By Ishpreet Khanuja
When OpenAI launched ChatGPT in late 2022, we witnessed the world shift almost overnight. The first global wave of accessible artificial intelligence had arrived, and with it, a sense that the boundaries of ideation and creation had permanently expanded.
But despite its ambitious promise of a simpler, less burdened world, the present day still seems more complex and crowded with doubt than ever before. Every headline claims a breakthrough. Every product calls itself ‘powered by AI’. The pace of progress is so fast that it is easy to mistake motion for meaningful change. For founders, that is the paradox of the moment: everything feels possible, yet nothing feels truly new. So, what does innovation look like in a world where everything, as we know it, is changing all at once?
In a recent Altitude Accelerator webinar, entrepreneur and former quantum physicist Amir Feizpour spearheaded an introduction into innovation with Large Language Models and Agentic AI. Feizpour is the founder, CEO and Chief Scientist at Aggregate Intellect, an AI venture studio focused on agentic workflow automation for business workflows. With a career that has its roots in quantum experiments at University of Toronto and University of Oxford to Natural Language Processing (NLP) at Royal Bank of Canada, he has spent the past decade at the intersection of research and application, building systems that connect reasoning, retrieval and automation into usable frameworks.
Understanding Agentic AI
Feizpour defines agents as software products that use AI to reason and act on behalf of a human user, not as a tool, but as a partner in digital or knowledge work.
“We throw agents as a term around,” he shares. “But it has a lot of layers of complexity around what it is, how it is set up and what kind of ingredients it needs.”
He further identifies two key components in the definition: autonomy, or the ability of a system to plan and act without continuous human supervision, and agency, in its capacity to communicate and collaborate effectively with human users.
This distinction separates agentic systems from current chatbots or prompt-driven tools. Traditional large language model interfaces wait for instructions and then respond. Agentic systems, by contrast, operate through control loops—plan, act, observe, and adapt—so that their behaviour evolves based on context and feedback.
The Future is Still Human
Feizpour emphasizes that the goal is not creating self-directed machines but building structured collaboration between human and computational reasoning. The more thoughtfully we design those loops, the more reliable and transparent the systems become.
Soon, he imagines, a founder is no longer just a person managing a team of dozens, but as a single individual surrounded by a variety of intelligent systems, with each agent handling specialized tasks like engineering, operations, outreach or analysis.
“I describe it as multi-agent, multiplayer systems,” he shares. “And I think that they will be the next frontier. Like today, you might have Claude as one of your agents. But John and Jane who work in your company should also be involved in this conversation as co-workers.
The Beginning of KnowledgeOps
Feizpour situates this shift within a broader history of how software systems have evolved. The advent of DevOps transformed software delivery by closing the gap between development and operations. Then came MLOps, which integrated the processes of training, testing, and deploying machine learning models into streamlined processes. Now, he argues that we are entering the era of KnowledgeOps and the operationalization of reasoning itself.
He notes, “In both scenarios (DevOps and MLOps), there was a gap: the process of observing data and deciding what to try next is still entirely manual. It’s a cognitively demanding task that requires domain knowledge, logic and reasoning. Tools like LLMs are now reducing that cognitive load. They provide powerful ways to brainstorm, interpret data, and identify the next experiments to run. I think we’re entering an era where the barrier between knowledge work and the operations that support it is dissolving—these functions are beginning to integrate.”
For founders, this represents a unique opportunity. The challenge is no longer building smarter models but building smarter systems around them that can justify their output, learn from feedback and work with real-world constraints.
Where We Stand Today
To see how far we have come since ChatGPT’s debut, Feizpour maps the last three years of product patterns. First came the text interface revolution and the novelty of a machine that could respond in natural language, an overnight success that had been years in the making. After realizing that, in theory, we could get better solutions in relevant informational ecosystems, we shifted towards RAG (retrieval-augmented generation), the method of grounding model responses in external data to curb AI hallucinations. What followed was a host of vector database companies, all promising to make generative AI enterprise-ready.
By 2024, the conversation had shifted again. We had realized that connecting a model to databases wasn’t the same as building a business-grade intelligent system. The true challenge wasn’t generation, but integration. How could we combine reasoning, retrieval, memory and verification into a single, secure and reliable process?
Feizpour suggests, “It is a grand vision of the future to build AI agent systems that can integrate seamlessly into the wild chaos of our world, with its unpredictable environments and actors. We want these agents to navigate that complexity gracefully and deliver on our objectives. But the reality is that we’re still at the beginning of that journey. Much of what we can build or use today must be heavily guard railed, controlled, and focused on a very specific set of workflows—for now, at least—until we develop systems sophisticated enough to truly deliver on that vision.”
Building Solutions That Work
When asked whether developing AI models for specific domain use is a good idea for solving existing problems in the field, Feizpour shares:
“My answer is: absolutely, with the caveat that the model itself doesn’t always need to be vertically aligned. I think the AI system, needs to be vertically aligned for a very specific and niche use case. The model might be generic, like a GPT model or Claude, but it is sitting inside a system design that is very specifc to the problem at hand.”
He contrasted the expensive path of training domain-specific reasoning models with a more accessible path for most founders: build AI systems that use pre-trained models, perhaps with fine-tuning and smart system design, targeted at a specific use case inside a vertical. Can Companies then get to market, collect data, and consider more well-informed model work? He concludes, “The answer is yes, but the path depends on how much money is in your bank account.”
To help founders reason where to build, Feizpour drew a map with three axes: task complexity (rising to the right), available process data (vertical), and established best practices/accuracy constraints (a secondary horizontal). He drew boundaries to show where current techniques perform: pretraining and decoding work well in the top area with rich process data and lower complexity, RPA augmented with LLMs expands what if-else can handle and verifiable domains like coding/math fuel reasoning models.
The broad ‘no man’s land’ in the middle, he argued, calls for ‘multimodal composite systems’ that are static at first, then increasingly trainable, mentioning DSP as an early indication of optimizing systems, not just models. His own team’s research, he claims, is very squarely focused on trainable composite systems that learn from user interactions.
The Conscious Builder
If there’s one theme that runs through Feizpour’s philosophy, it is intentional. “We are the architects of the future we are stepping into,” he reflects. “We can complain about corporate greed or governments, but we are the ones building the tools that shape what comes next.”
Awareness of responsibility and consequence is missing in much of today’s AI discourse. In a world driven by speed, the temptation is to treat innovation as an arms race: who ships faster, who scales first, who raises bigger rounds. The next generation of founders, Feizpour believes, will win not by building faster, but by building consciously.
In his venture studio, Feizpour uses AI to expand human capacity without eroding human judgment. To demonstrate what that looks like and how it helps founders think, Feizpour walked us through the mechanics behind his own workflow tooling.
He relies on command-line tools and keeps prompts as text files. AI tools like Claude Code can read these files and execute them as an agentic workflow. Much of his work processes, including operations, marketing, sales, engineering, and even receiving coaching is handled by these workflow files. He opened an ‘ideation framework’, as an example, where each step is clearly defined with its goal, input (‘read this text file as the prompt’), and output (‘write to this template file’).
As a live demo, Hessie Jones proposed a business problem: the time-consuming, error-prone work of monthly reporting across multiple systems (HubSpot, Salesforce, Excel) for compliance. Reframing it as a founder product, Feizpour then spoke to the workflow, feeding the agent natural speech via a transcriber. The system followed the steps he had defined and immediately responded with structured questions, e.g., clarifying which numbers are extracted, frequency, formatting constraints, and the “discovery story.”
As Jones explained the problem further (3–4 hours monthly, cross-references in Salesforce to verify client status, then spreadsheet compilation for Ministry reporting), the agent summarized the problem and asked focused follow-ups: What must the AI know? Are there constraints or possible hindrances? Who else suffers from the same pain? What would customers pay?
Feizpour highlighted a gap assessment step he bakes into these workflows, where the agent calls out vague phrasing. To push thinking beyond attachment to one idea, the agent presented two brief case stories and asked what might have gone wrong. Jones noted privacy/compliance concerns and customization burdens as her answers. The system recorded those and moved to reframing the problem statement, adding the validated risks. The workflow ended by prompting for a concrete, small experiment commitment.
The live demonstration was a powerful example of how an agent can coach a founder’s reasoning, amplify and ground their thought process, document assumptions and produce a testable plan without reasoning or a process that is hidden inside software the user can’t access or understand.
Reclaiming Our Own Agency
By the end of Feizpour’s session, the attendees weren’t just thinking about AI. They were thinking about themselves, and their own role as founders, designers and decision-makers. As we head towards an age of technology that no longer waits for permission, our choices have never been more consequential.
“What I’m encouraging you to do is to think about the implications of what you’re building, acknowledging the constraints like regulations and the capitalist world that you’re living in,” he urges. “Those are constraints that exist. We cannot deny them. We cannot be wishful and say, ‘Oh, it will work itself out.’ It will not. We have to build solutions that acknowledge those constraints, but are responsible. and meaningful in terms of the future that we would like to have.”
Progress is no longer measured by how much we can automate but by how precisely we can align autonomy and agency with human goals and prove it works. In that way, the founders of 2035 are not displaced by AI, but amplified by an exoskeleton powered by AI agents, able to explore more ideas, invalidate the wrong ones faster and bring the right ones to market with a level of traceability that earns trust.
The frontier, as Feizpour put it, will always move. The question is whether we are the ones chasing it or shaping it, in the way we build solutions.
Catch the full session with Amir Feizpour here to learn more about how founders today can design systems that reason, adapt and build with purpose.
We are accepting applications for the next Investor Readiness Program
If you are building something transformative and looking to accelerate your path to funding, Altitude Accelerator can help. Apply now.
