A history of agents
Side tracking into the rabbit hole
This one is not a practical post. I don’t have any AI tool tips, agents, or concepts for you. In this post, I’ll simply talk about the history of AI agents. Just to get context.
First off, “AI agent” is not a new term. The foundational ideas trace back to the 1950s and early computer science research. If you search for the early demos in computing back then, you’ll see some awesome and crazy work. Just watch this. And also this. To me, it’s still awesome 60+ years later.
John McCarthy and Oliver Selfridge at MIT were among the first to describe the concept of intelligent agents in the late 1950s and early 1960s, though they didn't specifically use the term "AI agent." Their work on programs that could sense, reason, and act in an environment laid important groundwork.
The first I’m aware of "agent" in an AI context is from the 1980s. Carl Hewitt's work on the Actor model in the 1970s was influential in the concept of computational agents that could act autonomously and communicate with each other. So the idea is OLD.
The specific phrase "AI agent" appeared with research interest in autonomous systems and distributed AI. The definition I found came from Stuart Russell and Peter Norvig in their 1995 textbook "Artificial Intelligence: A Modern Approach." It defines an agent as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment."
What “sensors” and “environment” are was left rather open.
Enter the Matrix
The 1999 movie was both inspiring and terrifying in many ways. The idea that AI Agents can decide and act faster than humans is something that gave pause at the time, but was still in the realm of sci-fi. At the time, there was still an AI winter. This is a term for a period of time with little funding of AI and little optimism.
Just a couple of years later though, that changed. Maybe the movies were an inspiration. In any case, the term “ai agents” made a comeback about a decade later.
Around 2015-2018, there was increased interest in AI agents in the context of reinforcement learning, particularly with DeepMind's work on game-playing agents like AlphaGo and AlphaZero. These were often described as "AI agents" that could learn and make decisions in complex environments.
Basically, these were gaming bots that were far more advanced than typical computer players in games. They kicked ass in Chess and Go. I’m more of a computer gamer, so I watched AlphaStar kick ass in Starcraft 2.
Resurgence
The recent revival of the term, however, came towards the end of 2023, in the context of "autonomous AI agents". These could perform tasks by breaking them down into steps and interacting with software tools. This was driven by several developments:
The release of Auto-GPT in early 2023, which was one of the first widely-discussed implementations of an autonomous AI agent that could pursue goals by breaking them into subtasks
Research papers and projects from companies like Anthropic ("Constitutional AI") and Google DeepMind (focusing on agent architectures) that explored AI systems with greater autonomy
Microsoft's introduction of plugins and tools for ChatGPT, which enabled the AI to interact with external software and complete more complex tasks
The meaning of "AI agent" in modern usage has evolved somewhat from its classical AI definitions. It still refers to systems that can perceive and act in an environment, there's now an emphasis on autonomous goal-seeking behavior, the ability to use tools and APIs, planning and reasoning, and memory and context retention.
The environment mentioned by Russel and Norvig is our current internet and software environments.
For the superfans
I’m not an ai agent expert. More of a SuperFan. If you’re looking for the technical experts that have dealt with this topic for decades, you’ll find them here:
IFAAMAS - The International Foundation for Autonomous Agents and Multiagent Systems is a non-profit organization whose purpose is to promote science and technology on autonomous agents and multiagent systems.
AAMAS - the International Conference on Autonomous Agents and Multiagent Systems is where those scientists go to share their research. There is also a journal.
MultiAgent Systems - a seminal book by Gerhard Weiss. Gerhard is a prominent figure in the field of multiagent systems and distributed artificial intelligence. He’s been writing about this stuff since the mid 90s.
You can also follow Michael Wooldridge’s work - he’s a CS professor in the UK and wrote a couple of books on the topic. The latest one is very accessible and takes you through the history in detail.
Happy Reading!
I know this one was historic & more technical. Let me know if you prefer either more A) applicable & business oriented posts, B) historical & philosophical, C) mix of both.