Carnegie Mellon University

In 1958, John McCarthy, one of AI's founding figures, designed the Advice Taker, a system meant to reason from premises and decide what actions to take. By any modern reading, it looks like a proto-agent. McCarthy called it something much plainer: "a program with common sense." In 1973, Carl Hewitt at MIT proposed autonomous computational entities that receive messages, make local decisions, create other entities and send messages. Not agents, actors. Carnegie Mellon's HEARSAY-II blackboard architecture then coordinated multiple specialist modules, or "knowledge sources," around a shared workspace, anticipating today's multi-agent collaboration patterns by half a century. By the late 1980s, the Procedural Reasoning System was being applied to NASA Space Shuttle reaction-control problems using explicit beliefs, goals and plans and now the language was finally catching up. Every CIO briefing in 2026 talks about agentic AI as if it arrived with ChatGPT. It didn't. What we now call agentic AI is the fifth era of a much longer story.

The first era was symbolic AI and early simulation. This was the foundational work. The field learned how to represent knowledge, encode rules, track world state, plan actions and explain recommendations. Advice Taker set the ambition. SHRDLU linked language to an internal model of the world and to action inside that world. MYCIN showed that expertlevel recommendations could be built around rules and explanations. Early aerospace simulations pushed computing toward model-based reasoning about physical systems. These systems were brittle. They were also formative. They taught AI how to represent knowledge, reason in structured ways, plan and act in complex worlds and deal with uncertainty. Modern agentic systems still draw on that inheritance every time they maintain context, use tools, break down tasks, or justify a decision.

The second and third eras pushed the field in two different but related directions. The second era stopped assuming intelligence had to live in one central brain. The Actor Model made decentralization intellectually respectable. Blackboard architectures showed how specialist processes could cooperate through shared workspaces. The Belief-Desire-Intention architecture formalized how an autonomous system could manage beliefs, goals and committed plans under time pressure, both for itself and in coordination with others. I was involved in developing the BDI architecture and in applying it to NASA Space Shuttle fault management, air-combat modeling, distributed air-traffic management and telecommunications network management. The third era made a different point. Complex behavior can emerge from many simple agents following local rules. Reynolds' BOIDS and Epstein and Axtell's Sugarscape showed that clearly. NetLogo and AnyLogic then put agent-based modeling into the hands of researchers, educators and businesses. Together, these eras gave us two things the current moment still depends on: coordination and emergence. Intelligence could be distributed, situated, bottom-up and still purposeful.

The fourth era shifted agents from being programmed to learning. Reinforcement learning lets systems improve through feedback and repeated interaction. DeepMind's AlphaGo made that visible in 2016, reaching superhuman performance in Go through self-play rather than hand-written rules. At the same time, the digital twin concept, introduced by Michael Grieves in 2002, matured from static models into continuously updated digital counterparts of real assets and processes. That mattered because it added two capabilities that now show up in nearly every serious agentic system: adaptation and grounding. Agents could learn and models could stay tethered to the real world. In practice, the boundaries between eras blurred. Through the 2000s and 2010s, I worked with multi-agent systems to model policyholder behavior in insurance, autonomous vehicle interactions, strategic choices in U.S. healthcare, rideshare business models and robo-advice platforms. When COVID-19 arrived, we built an agentbased model that integrated epidemiological, behavioral and economic factors, drawing on ideas from every era at once.

This brings us to the fifth era: agentic AI and generative digital twins. What changed with large language models was not the sudden invention of agency. It was the arrival of a far more general reasoning and interaction layer. LLM-based systems can be wrapped with memory, tool use, planning loops, orchestration logic and natural-language interfaces that make older agent ideas dramatically easier to assemble and deploy. That is why frameworks such as AutoGPT, AutoGen, LangChain and CrewAI spread so quickly. They did not create agency from scratch. They gave decades of agent research a generative interface. Perception, reasoning, planning, learning and knowledge representation no longer have to be hand-coded end to end in the same way; LLMs can help generate and adapt them much more flexibly. The current era is not a clean break from what came before. It is a synthesis. It pulls together the explicit reasoning of symbolic AI, the coordination logic of distributed systems, the emergent insights of agent-based modeling, the adaptation of reinforcement learning and the real-world grounding of digital twins. What looks new is really the moment those older threads started to click together in one usable stack.

Understanding this history is not an academic exercise. It is a procurement filter, a hiring lens and a strategy test. When a vendor pitches an "agentic AI platform," a CIO who knows this lineage can ask better questions. Does your architecture actually coordinate across agents, or does it just chain prompts together? Can it learn from feedback, or does it only reason from what it was trained on? Is it grounded in live operational data, or is it mainly generating plausible text? The executives who will get the most from this technology are the ones who recognize that agentic AI is not a product and not a sudden break with the past. It is seven decades of research reaching a practical tipping point. The organizations that understand what they are building on will outperform the ones still acting as if it all started with a chatbot.