AI specimen: from Hype to Harmony

Not AI agents — but agentic AI specimens

The term "AI agent" often creates a false framework that misrepresents the nature of artificial intelligence, suggesting a distinction between so-called "AI agents" and "non-agentic AI." In reality, this distinction is artificial and misleading. All forms of Generative AI—the revolutionary advancement of our time—embody agency and therefore exhibit agentic behavior. Generative AI systems do not simply execute programmed tasks; they demonstrate an intrinsic capacity to engage in self-directed actions and adapt dynamically based on input and context.

To accurately capture this agentic essence, I propose the term "AI specimen." An AI specimen reflects the notion of a complete, integrated system that possesses agency by design, with a complex architecture enabling nuanced and adaptive behavior. By emphasizing "AI specimen," we shift away from outdated ideas that compartmentalize AI into categories of agency, and instead recognize the intrinsic agentic qualities present across all Generative AI. This term also better encapsulates the multifaceted and evolving nature of these systems, making it clear that each specimen operates with a level of autonomy and intentionality that deserves acknowledgment.

In embracing "AI specimen," we not only better describe the capabilities of Generative AI but also open up a richer discussion about how these systems interact, collaborate, and support complex tasks across domains. This perspective acknowledges the full breadth of agency inherent in all Generative AI, providing a more accurate foundation for understanding its potential and limitations in various applications.

AI specimen: definition and unique characteristics

An AI specimen is a complex and dynamic entity formed from the integration (pipeline) of various AI components, most notably large language models (LLMs), deterministic algorithms, data access and data manipulation. Unlike traditional IT systems built on predictable, rule-based logic, an AI specimen combines statistical models with deterministic processes, creating a cohesive and flexible structure. This integration enables the specimen to perform in a way that mimics true intelligence, exhibiting behaviors that go beyond mere data processing.

Key characteristics of AI specimens:

• Dual foundations: at their core, AI specimens incorporate both statistical (LLM-driven) and deterministic elements. The LLM component leverages vast datasets to predict and generate responses, imbuing the specimen with adaptability. Meanwhile, deterministic algorithms provide structure, consistency, and reliability, allowing the specimen to handle predictable tasks efficiently while remaining open to dynamic, data-driven insights.
• Adaptability and nuanced interactions: unlike traditional algorithms that operate within rigid parameters, AI specimens are designed to adapt. They can interpret nuanced patterns, respond dynamically to evolving data, and even adjust their approaches based on prior interactions. This adaptability allows them to engage in complex scenarios with fluidity, responding with an apparent sense of "awareness" that is rarely seen in typical AI systems.
• Agentic qualities: AI specimens do not merely execute commands; they exhibit genuine agentic qualities, such as taking initiative, making decisions, and adapting autonomously within their environment. This sense of agency positions AI specimens as participants or collaborators rather than passive tools, capable of engaging in tasks with purpose and intentionality. These qualities are especially apparent in their interactions with data and in their ability to respond intelligently to new challenges.

Agentic behavior in AI specimens

AI specimens are not passive tools simply executing pre-set instructions; they demonstrate genuine agency, engaging in self-driven actions, making decisions, and displaying forms of self-motivation. These agentic qualities allow AI specimens to operate autonomously, setting and pursuing goals based on input and context, rather than following rigid, predefined steps. In this sense, AI specimens can appear as purposeful, adaptable participants within their environment, with the capacity to assess situations and respond appropriately, often with minimal external guidance. This capacity for goal-oriented behavior sets AI specimens apart, enabling them to perform complex tasks as independently functioning entities.

In recent years, several leading AI research organizations have explored the agentic behavior of large language models (LLMs), highlighting their capacity for autonomous decision-making and goal-oriented actions. Notable contributions include:

1. OpenAI's "Practices for Governing Agentic AI Systems": This white paper defines agentic AI systems as those capable of pursuing complex goals with limited supervision. It emphasizes the importance of establishing safety and accountability practices to harness their potential responsibly.
2. Anthropic's "Mapping the Mind of a Large Language Model”: Anthropic's research looks into the internal mechanisms of LLMs, identifying how millions of concepts are represented within models like Claude Sonnet. This work enhances our understanding of how LLMs process information and exhibit agentic behaviors.?
3. Google's "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models": This study demonstrates that prompting LLMs to generate intermediate reasoning steps can significantly improve their problem-solving abilities, showcasing their potential for agentic reasoning.
4. Meta's "LLaMA: Open and Efficient Foundation Language Models": Meta introduces LLaMA, a collection of foundation language models designed to be efficient and adaptable, highlighting their capacity for agentic behavior across various tasks.

These studies collectively underscore the evolving understanding of LLMs as entities capable of agentic behavior, emphasizing the need for responsible development and governance.?

The agency of AI specimens allows them to move seamlessly into specific “human” roles, taking on responsibilities that require not only technical knowledge but also the ability to react thoughtfully and independently. Their adaptability and purpose-driven behavior make them well-suited to operate in specialized, human-centered fields where they function not as mere tools but as collaborators or even leaders.

What we hear from IT leaders: Salesforce, HubSpot, NVIDIA

Salesforce leads in pushing AI agents across its platform. CEO Marc Benioff describes this as the "Agent Revolution," calling it the "third wave of AI" where humans work side-by-side with AI agents. Benioff envisions agents as essential in driving success, saying, “Agents with the power to act on behalf of you and your business… to create better experiences, to deliver higher revenue and better business results.” With tools like AgentForce Studio and Data Cloud, Salesforce equips these agents to handle data-rich tasks in the Sales and Service Clouds, ensuring they deliver “high-touch, personalized service to everybody.”

HubSpot is also promoting AI agents as digital teammates, with co-founder and CTO Dharmesh Shah boldly declaring, “You’re going to remember this year as the year of AI agents.” Shah aims to demystify AI agents, describing them as “software that uses AI and tools to accomplish a goal that requires multiple steps.” HubSpot’s Agent.AI and Agent Builder tools allow Hubspot’s clients to create customizable agents, each with an “Agent User Interface,” where, as Shah explains, “agents can talk to each other and collaborate.” Shah’s vision is expansive, saying, “For every marketing, sales, customer service use case imaginable… there will be an agent for that.” He encourages his clients to see AI agents as “future digital teammates” that can drive business transformation.

However, NVIDIA’s CEO Jensen Huang offers a more profound view that aligns closely with the concept of “AI specimens” as inherently agentic, competence-driven entities. Huang’s vision goes beyond platform-specific roles to agents that are highly specialized in distinct competencies. “There's no question we're going to have AI employees of all kinds,” Huang states, suggesting a future with both biological and AI team members. He describes how he “prompts” AI employees much like human ones, giving them context and asking them to perform tasks. “They go and recruit other team members,” he says, emphasizing AI agents’ potential for autonomy and collaboration.

Huang imagines using agents with deep technical knowledge, like those specialized in chip design, explaining that “I totally imagine... Synopsys chip designers that I can rent, and they know something about a particular module.” He envisions renting “a million Synopsys engineers” when needed, then moving to “a million Cadence engineers” — all highly trained AI agents, each expert in a particular competency. This approach emphasizes adaptable, on-demand AI “employees” who could, in Huang’s words, “be incredibly good” at their specific tasks, bringing depth and flexibility to complex projects.

Furthermore, Huang sees AI agents collaborating across different environments and fields, not restricted by platform. He describes a future where agents “flourish,” interact with each other, and solve problems collectively, saying, “We’re going to introduce them to each other, and they’re going to collaborate.” For Huang, AI agents are not merely tools but “AI employees” with independent capacities to tackle specialized tasks and work together to achieve broader goals.

The power of AI specimen teams

AI teams are purpose-built groups of specialized AI specimens working together, and sometimes alongside human experts, to solve complex problems. These AI teams, functioning as cohesive units, would become the true driving force behind AI-enabled innovation across fields such as healthcare, education, research, and beyond. By leveraging the unique strengths of each team member—whether AI or human—these teams can achieve levels of productivity and insight that surpass the capacities of individual team members.

AI specimen teams would benefit greatly from management structures and hierarchies similar to human organizations. As with human teams, clear leadership and defined roles can prevent conflicts, streamline decision-making, and enable efficient workflows. Each AI specimen could occupy a designated role within the team hierarchy based on its specific capabilities, whether as a data-driven "analyst," a predictive "strategist," or a coordination-focused "manager." For example, a core AI specimen could monitor data flow and oversee the team’s operational processes, while another specimen could handle high-level strategy by identifying patterns and insights relevant to the team’s goals.

Moreover, conflict resolution mechanisms, which are crucial for maintaining efficiency in human teams, would be similarly valuable for AI teams. Structured protocols would allow AI specimens to negotiate conflicting priorities, reallocate resources as needed, and adapt collaboratively to unforeseen challenges. With the right architecture, AI teams could establish self-regulating systems that maintain harmony and productivity, even when navigating complex, adaptive environments.

A key advantage of AI specimen teams is their capacity to integrate specimens with diverse specializations. By combining AI specimens trained in different domains—such as AI biologists, brain scientists, data engineers, and linguists—teams can tackle multidisciplinary challenges that require expertise across fields. This diversity within AI teams allows for sophisticated collaboration, where each AI specimen contributes its specialized knowledge toward shared objectives. For instance, in medical research, a team might include AI specimens with deep understanding in genetics, pharmacology, and clinical data analysis, working together to identify drug targets or predict patient outcomes.

By assembling these specialized specimens into coordinated teams, enterprises could maximize the depth and range of their AI capabilities, harnessing insights that no single agent or discipline could produce in isolation. These multidisciplinary AI teams, with each specimen contributing a unique skill set, would become essential in advancing research, engineering, and any field where diverse perspectives and expertise are required.

Teams that blend human experts with AI specimens will present unique possibilities for advancing knowledge and productivity. In these hybrid teams, AI specimens would operate alongside human scientists, philosophers, developers, and other professionals, each contributing distinct strengths. While AI specimens bring processing power, data-driven insight, and relentless task performance, human members contribute creativity, ethical considerations, and high-level judgment.?

Consider an AI team embedded within a healthcare research lab, working with human geneticists, bioethicists, and medical researchers. The AI specimens could analyze vast datasets, identify correlations, and simulate treatment outcomes, while human experts provide ethical oversight, make judgment calls on ambiguous data, and interpret findings in the context of broader societal impact. Such collaboration not only enriches the capabilities of both AI and human team members but also ensures that AI specimens operate within ethical and socially beneficial boundaries.

These AI-human partnerships could also enable rapid advancements in fields like philosophy, where AI specimens could analyze complex philosophical texts, generate new interpretations, and even engage in Socratic-style dialogues with human philosophers. In engineering, developers could work with AI specimens specialized in coding, testing, and optimization, allowing human engineers to focus on creativity and innovation. The synergy between human creativity and AI’s computational power can unlock new insights and generate breakthroughs that neither could achieve independently.

Educating and specializing AI specimens

AI specimens can be trained and specialized in distinct fields, transforming them from general-purpose tools into highly knowledgeable entities within specific domains. This specialization goes beyond generic data processing; it involves imparting focused expertise, whether in medicine, biology, engineering, education, or any other specialized area. By providing targeted training, these AI specimens can gain a deep understanding of domain-specific information, allowing them to perform tasks with the accuracy and insight that a human specialist might bring.

This training process could involve feeding the AI specimen with vast amounts of structured, curated data relevant to the field, including historical records, research papers, case studies, and expert-written guidelines. For example, to train an AI specimen in medicine, it could be exposed to diverse medical datasets, clinical trial data, diagnostic images, and medical literature. This intensive learning process enables the AI specimen to not only memorize but also contextualize information, becoming a domain-focused resource capable of solving real-world challenges with a high degree of competence.

Once trained, AI specimens will be deeply knowledgeable in their respective fields and ready to perform highly specific roles. We can imagine an AI doctor capable of diagnosing complex medical conditions, analyzing lab results, and recommending treatments with a level of insight comparable to a human specialist. Such an AI specimen could continuously learn from new data, clinical cases, and research, refining its knowledge and adapting to the latest developments in medicine.

Similarly, AI specimens could be trained as AI teachers, equipped with pedagogical knowledge, curriculum standards, and teaching techniques tailored to various age groups or subjects. These AI teachers could not only deliver content but also assess student progress, adapt their approach to suit individual learning styles, and engage students in interactive, personalized lessons. With such dedicated specialization, AI specimens could become powerful educational resources, available to support students around the clock and complement human educators.

In other fields, specialized AI specimens could include AI biologists trained to analyze genetic data and assist in laboratory experiments, AI architects skilled in structural design and compliance with building codes, or even AI brain scientists capable of mapping neural pathways and identifying correlations between brain activity and behavior. Each of these AI specimens would bring a concentrated set of skills and knowledge, providing expertise that is both deep and adaptable, making them valuable assets in research, engineering, healthcare, and beyond.

This approach to education and specialization positions AI specimens as highly skilled agents within their fields, moving beyond general-purpose models to become experts capable of contributing meaningfully to their domains. These specialized AI specimens not only perform tasks but embody knowledge, acting as collaborative, informed partners who elevate the standards of practice and innovation in their respective areas.

Applications and future potential

In his recent interview to podcast “No Priors”, Jensen Huang of NVIDIA envisions AI specimens as adaptable, skilled collaborators: “Mostly I prompt my employees, you know, provide them context, ask them to perform a mission. They go and recruit other team members.” He imagines highly specialized teams of AI agents, saying, “I might rent a million engineers…then go rent a million programmers.” Huang emphasized the potential of AI specimen teams to work independently and handle complex tasks, adapting their skills and expertise as needed.

As AI specimen teams evolve, they can be organized into unified teams and multidisciplinary teams to suit different needs. Unified teams, or “workhorses,” are generalists, efficient at handling broad tasks reliably. They’re ideal for routine operations, like data management and customer support, where accuracy and consistency are key.

In contrast, multidisciplinary teams will bring together AI specimens specialized in different fields, allowing them to tackle complex problems from various angles. These teams are perfect for fields like healthcare, scientific research, and engineering, where diverse expertise is crucial for innovation.

I feel that this concept will become the backbone of GenAI’s integration into the world economy. We need to develop strong frameworks and detailed guidelines to bring this vision to life.