Creating Collaborative Systems: Amir Feizpour on the Evolution of Agentic AI Past, Present, and Future

The launch of ChatGPT by OpenAI in late 2022 marked a significant turning point in the accessibility of artificial intelligence, expanding the potential for ideation and creativity globally. However, despite this exciting development, today's landscape remains complex, filled with conflicting narratives about innovation. Founders find themselves in a paradox where although everything appears possible, nothing seems fundamentally novel.

In a recent Altitude Accelerator webinar, Amir Feizpour, an entrepreneur and former quantum physicist, introduced the concept of innovation using Large Language Models (LLMs) and Agentic AI. He is the CEO and Chief Scientist at Aggregate Intellect, an AI venture studio focused on automating business workflows. Feizpour's background spans quantum physics and Natural Language Processing, positioning him favorably at the intersection of research and practical application.

Feizpour explains that agentic systems differentiate from traditional chatbots. These software products utilize AI to operate more as collaborators rather than mere tools, embodying the capabilities of autonomy and agency. Autonomy allows systems to act without continuous human oversight, while agency enables effective communication with human users. This is where agentic systems differ from conventional LLM interfaces that only respond when prompted.

The emphasis, according to Feizpour, is on creating structured collaborations between humans and AI, rather than merely developing self-directed machines. The vision for the future includes environments where founders interact with diverse intelligent systems handling distinct operational tasks, effectively transforming team dynamics.

Feizpour aligns this shift with historical developments in software systems, suggesting we are entering an era of KnowledgeOps—where reasoning itself is being operationalized. He highlights the limitations of past innovations like DevOps and MLOps in that they still required significant human cognitive input for data interpretation and decision-making. Now, tools like LLMs can alleviate this cognitive burden, seamlessly integrating knowledge work with operational execution.

This shift poses a unique opportunity for founders: the focus now is on creating intelligent systems capable of self-justifying their outputs and adapting based on real-world constraints. Reflecting on recent developments since ChatGPT's launch, Feizpour describes an evolution where advances led from simple text interfaces to more sophisticated Retrieval-Augmented Generation (RAG) techniques aimed at grounding model outputs in external data. However, the challenge remains shifting from generation to integration, as companies must find ways to combine reasoning, retrieval, memory, and verification into secure processes.

Feizpour envisions a future that integrates AI systems into the complexities of real-world environments, navigating unpredictability while achieving set objectives. However, he warns that we are still in the early stages, where existing systems often need strict boundaries and focus on specific workflows.

Regarding AI models for specific domains, Feizpour advocates for this under the condition that the system design is well-tailored to niche applications, even if a generic model powers it. He emphasizes that while training domain-specific models can be costly, leveraging pre-trained models adjusted for specific use cases is a more feasible path for most founders.

To guide founders, Feizpour presents a framework along three dimensions: task complexity, available process data, and established best practices/accuracy constraints. This mapping aids in identifying where current AI techniques excel and where new opportunities lie, paving the way for innovative solutions in AI-driven businesses. Ultimately, while the landscape is vibrant and evolving, the path to realizing the full potential of these technologies is still forming.