From One Neuron to AI Agents: The 80-Year Story of Agentic AI

An AI agent that schedules meetings, researches markets or fixes bugs can trace its lineage to a 1943 paper on brain cells. While we marvel at what AI can do today, the surprising foundation of agentic AI is history. These foundations span eight decades, with each era addressing one fundamental question: Can a machine compute, learn, focus, understand and act?

Here's the journey, one question at a time.

Can a machine compute? The artificial neuron (1943)

In 1943, neuroscientist Warren McCulloch and logician Walter Pitts wrote the first mathematical model of a brain cell. Their "neuron" does one simple thing. It adds up incoming signals and fires only when the total crosses a threshold. That rule is the ancestor of the activation function in every modern AI model.

Think of it as a dimmer switch. Many small signals feed in. The light comes on only once they pass a set point. Wire billions of these switches together, and you have the basic building block of today's largest models.

McCulloch and Pitts proved that networks of these simple units could, in principle, compute anything logic can describe. They did it two years before the first electronic computer, ENIAC, was switched on. It was pure math, and it became a blueprint.

Can it learn? Perceptrons, backpropagation and deep learning (1958–2012)

A neuron that can't learn is just a fixed calculator. In 1958, Frank Rosenblatt changed that with the perceptron, the first artificial neuron that learned from examples. Each time it got an answer wrong, it nudged its internal "weights" and tried again. It first ran as software on an IBM 704 in 1957, the first time modern AI ran on a digital computer.

Then came the limit. A single-layer perceptron can only draw straight-line boundaries. It couldn't learn even a simple either/or pattern, known as XOR. In 1969, Marvin Minsky and Seymour Papert proved it. Add hype that had run far ahead of reality and computers far too weak for the job, and funding collapsed. The first AI winter set in by the mid-1970s.

The way out was clear on paper: stack many layers. Nobody knew how to train them. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams popularized the answer: backpropagation. When the network gets something wrong, trace the error backward through the layers and let every neuron adjust its share of the blame.

Think of it as tuning an orchestra. When a chord sounds off, the conductor doesn't retune every instrument at random. The error is traced back section by section, and each musician adjusts in proportion to how far off they were.

Backpropagation is still the algorithm that trains virtually every AI model in use today. Hinton, Yann LeCun and Yoshua Bengio, the researchers who kept the field alive through lean years, shared the 2018 Turing Award for it.

One more turning point came in 2012. A deep neural network called AlexNet, trained on two gaming GPUs, won the ImageNet photo-recognition contest by a landslide. It cut the error rate from about 26% to about 15%. The ideas were decades old. What changed was the fuel: the internet finally supplied enough data, and gaming chips supplied enough computing power. Neural networks became the dominant approach to AI almost overnight.

Can it focus? Vectors, attention and the transformer (2013–2017)

Computers can't do math on words. So the next breakthrough was a translation layer between language and numbers.

Words become vectors (2013). Text is split into small pieces called tokens, and each token is mapped to a long list of numbers called a vector. A research team at Google showed with word2vec that these vectors can be learned so that meaning becomes geometry. Similar words land near each other. Directions carry meaning. The famous example: take the vector for "king," subtract "man," add "woman," and you land near "queen."

Think of it as map coordinates for meaning. Every word gets a location, and the distances and directions between words say something real.

Attention (2014). Until then, language models were recurrent neural networks, or RNNs. They read one word at a time and carried a running "memory" that faded over long sentences. Translation was the hardest test, because the model had to squeeze a whole sentence into one fixed-size summary first.

Dzmitry Bahdanau and colleagues added a fix: attention. For each word it produced, the model could glance back at every input word and weigh which ones mattered right now.

Think of it as a translator who keeps glancing back at the original page instead of working from a fading memory.

The transformer (2017). Then Google researchers asked a bold question: what if attention is all you need? Their paper, "Attention Is All You Need," threw away the RNN and built the whole architecture from attention. Every word looks at every other word at once, instead of passing a note down a long line of people.

That design had a second advantage that mattered just as much. Processing every word simultaneously suits GPUs perfectly. Training on enormous datasets suddenly became practical. Nearly every leading AI model today, for text, images, code and audio, is a transformer.

Can it understand? Large language models and ChatGPT (2018–2022)

With the transformer in hand, researchers tried something deceptively simple. Train a model on one goal, predict the next word, across a huge slice of the internet.

The results surprised everyone. OpenAI's GPT-3 (2020), with 175 billion parameters, could write, translate and even do basic coding. Nobody had explicitly taught it any of those skills. Researchers also found "scaling laws": more data, more computing power and a bigger model produce predictably better results. Scale itself became a strategy.

Think of it as autocomplete that has read most of the internet. Predict the next word well enough, across enough of human writing, and something that looks a lot like understanding begins to emerge.

But a raw language model completes text. It doesn't follow instructions or hold a helpful conversation. The fix was learning from feedback. In a technique called reinforcement learning from human feedback (RLHF), people rank the model's answers, and the model learns to give the answers people prefer.

In November 2022, ChatGPT put this in everyone's hands. It reportedly reached around 100 million users in two months, the fastest adoption of any consumer app to that point. AI became usable by anyone with a keyboard. No coding required.

Can it act? Reasoning, tools and agents (2023 and beyond)

A chatbot can talk. An agent can do. Two advances closed the gap.

Reasoning. Researchers found that simply asking a model to think step by step, known as chain-of-thought prompting, dramatically improved its reasoning on hard problems.

Tool use. Models learned to call calculators, search engines, databases and code, through capabilities such as function calling. Frameworks like ReAct let them alternate between thinking and doing: reason, act, observe the result, adjust, repeat.

Think of it as a brilliant consultant locked in a room who finally gets a phone, a laptop and permission to make calls. Same brain. Suddenly able to check facts, run numbers and get things done.

Put those pieces together and you have agentic AI: systems that plan multi-step tasks, act through tools, check their own results, retry when something fails and remember what happened. Open standards, such as the Model Context Protocol introduced in 2024, give models a universal plug for connecting to tools and data. Multi-agent teams now split complex work among specialists.

Think of it as the move from an assistant who answers questions to a coworker you can delegate a job to. You describe the outcome. The agent plans the steps, does the work, checks the results and comes back when it's done.

Three lessons from 80 years of AI

1. Nothing here was sudden. Today's agents are the product of decades of ideas that finally met enough data, computing power and connective tissue. When a technology looks like an overnight success, look for the long runway.

2. Winters follow hype, not ideas. The perceptron's limits were real, and so was the funding collapse. But the underlying ideas survived and came back stronger. Separate what a technology can do today from what it will do tomorrow, and plan for both.

3. Breakthroughs often start as patches. Attention began as a fix for translation models. Learning from feedback began as an alignment tweak. Chain-of-thought began as a prompting trick. Each one became the foundation of the next era.

What this means for your organization

Someone who knows the story has a mental model. They understand that AI learns patterns from data, became fluent by predicting the next word at enormous scale and gained the power to act through tools. So they stop treating AI as magic or as a threat. They start asking better questions. What was this trained on? What can it check? What happens when it's wrong?

The same history keeps expectations honest. AI is genuinely strong at language, code and routine multi-step work. It can also be confidently wrong, it knows only what it was trained on and what you give it, and agents can compound small errors. High-stakes actions need guardrails and human review.

The organizations that succeed with AI won't be the ones that believe the most hype or the most doubt. They'll be the ones whose people understand the technology well enough to use it wisely. 

Key references

• McCulloch, W. & Pitts, W. (1943). "A logical calculus of the ideas immanent in nervous activity." Bulletin of Mathematical Biophysics, 5, 115–133.
• Rosenblatt, F. (1958). "The perceptron: a probabilistic model for information storage and organization in the brain." Psychological Review, 65(6), 386–408.
• Rumelhart, D., Hinton, G. & Williams, R. (1986). "Learning representations by back-propagating errors." Nature, 323, 533–536.
• Krizhevsky, A., Sutskever, I. & Hinton, G. (2012). "ImageNet classification with deep convolutional neural networks." NeurIPS 2012.
• Mikolov, T. et al. (2013). "Efficient estimation of word representations in vector space." arXiv:1301.3781.
• Bahdanau, D., Cho, K. & Bengio, Y. (2015). "Neural machine translation by jointly learning to align and translate." ICLR 2015.
• Vaswani, A. et al. (2017). "Attention is all you need." NeurIPS 2017.
• Brown, T. et al. (2020). "Language models are few-shot learners." NeurIPS 2020.
• Kaplan, J. et al. (2020). "Scaling laws for neural language models." arXiv:2001.08361.
• Ouyang, L. et al. (2022). "Training language models to follow instructions with human feedback." NeurIPS 2022.
• Wei, J. et al. (2022). "Chain-of-thought prompting elicits reasoning in large language models." NeurIPS 2022.
• Yao, S. et al. (2023). "ReAct: synergizing reasoning and acting in language models." ICLR 2023.
• Anthropic (2024). "Introducing the Model Context Protocol."