Just when we thought we had a grip on all things AI, something new heaves into view. Here’s your introduction to agentic AI.
There’s always something new out of Africa.
Or so goes the old saying, which has always irritated me: it implies that there’s something special and mysterious about the continent (and things that are “special” get treated differently).
I looked it up, though, and apparently all it reflects is an old myth.
According to an Oxford Academic Tumblr post (always something new out of Oxford? A Tumblr?), it’s a proverbial expression that’s a translation of the Latin ex Africa semper aliquid novi. “The immediate source of the saying is a passage in the Natural History of the Roman scholar Pliny the Elder. Explaining the number of African animals by hybridization (for example, lions breeding with leopards), Pliny explains that this is what gave rise to what he calls a common Greek saying that ‘Africa always brings forth something new’.”
As XenophonTheAthenian says in a Reddit thread: “He seems to be under the impression that the various species of big cats are produced when members of different species interbreed.” In essence, the new things out of Africa were just animals that seemed weird to people in the ancient classical world.
Why are you reading this foray into ancient Rome and proverbial English? Because it’s tempting to take that old saying and adapt it: “There’s always something new in the world of AI.”
That something new is currently agentic AI. So what is this, what does it do, and what’s in it for the everyday user of a computer or cellphone?
What is agentic AI?
First, let’s go back to basics. For the last several years, we’ve all been playing around with generative AI tools like ChatGPT, which are based on Large Language Models (LLMs).
Generative AI is, as the name implies, a form of artificial intelligence, or the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings (for example, your phone using your face as an unlocking method). In November 2022, a demo of ChatGPT showed what the underlying LLM could do: give answers to prompts by analysing massive datasets of language, allowing it to recognise and interpret human language or other complex data.
With various additions and refinements, that kind of AI has been the subject of much hype over the last while. Even so, it is still essentially in the realm of a human being asking a machine to do a simple or complex task (or a series of them) and then using the result to complete a task in the real world. An example: looking at the contents of the fridge and asking ChatGPT or Claude or Gemini to come up with a recipe, which the human then shops for and cooks.
One step up from that: an AI agent
But things are moving fast, and these tools are changing. Enter AI agents.
As defined in a long and detailed article by Moveworks, a company which offers agentic AI solutions to companies, an AI agent is a ”software programme designed to understand its environment, process information, and take actions to achieve specific goals”.
The key bit of that is “take actions” – these agents are designed to operate independently. That might be something like an automated scheduling assistant booking meetings, or the AI agent that’s connected to your smart fridge might be able to alert you when the milk is running low.
Agents can work in teams: in customer service, one AI agent might process language, another might search knowledge bases, and a third could handle routing of customer requests, all working together to solve customer issues.
Put it all together and you get agentic AI
Agentic AI is based on AI agents, but it is one step up. Here, we get a system that can learn and reason, make autonomous decisions and take goal-driven actions. AI agents generally focus on single tasks while agentic AI uses multiple agents to handle complex workflows. Agentic AI could plan a dinner party, including planning the menu,, sourcing the recipes, shopping for ingredients and surveying the guests for calendar availability and dietary restrictions (along the way collaborating with the guests’ AI agents to understand those calendar and dietary constraints).
If that sounds scary to you, it should. There are risks!
Of course there are risks
The discerning reader will have wondered how it is that agentic AI would have done the shopping for the dinner party. You would have had to give it access to your retailer’s online shopping account to do that – and to invite the guests you would have had to give it access to your email account and their email addresses. I’m not sure I would want an AI system to be doing either of those two things.
I assume most people are many steps away from planning dinner parties with agentic AI, but here’s a brief list of possible problems:
Security threats: Agentic AI brings with it things with names like prompt injection, data leaks, model tampering, and data poisoning. Broadly, these are ways in which malicious commands can be inserted into the agentic process.
Accountability gap: When an autonomous AI system makes a harmful decision, determining who is responsible – the system, its developers, or the organisation that deployed it – becomes incredibly difficult.
Bias and discrimination: AI models learn from data, and if this data is non-diverse or unrepresentative, the models can perpetuate existing inequalities (think agentic AI that is used in surveillance where facial recognition can have higher error rates for certain demographics).
Ways in which agentic AI might filter down into everyday life
The dinner party example above is a good one. Here are some other suggestions, culled from a variety of sources that I fed into NotebookLM (the notebook is here):
Selling things: An AI agent can take a list of unused household items, write optimised product listings, set prices, and post them on online platforms, all without manual input.
Real estate research: A potential homebuyer could use agentic AI to gather comprehensive data on crime rates, schools and transport options and then have it summarise the pros and cons to aid in home selection.
General digital assistants: Users can have their own bots or AI agents that help them sift through emails, prepare meeting agendas, send calendar invites, or even contact individuals to escalate a complaint.
Personal health assistants: An agentic AI could act as a personal health assistant, monitoring data from wearable devices, adjusting treatment plans based on real-time health indicators, scheduling appointments, and predicting potential health issues before they become serious.
Where to from here?
This agentic AI thing is not new – but it’s appearing more and more in all the feeds that I follow and so I thought an introduction to the topic might be valuable.
I’m still in the beginning phases of learning about agentic AI, so for now I have no simple recommendations for ways to try this yourself, or views on how all the risks will play out.
I do have some thoughts though. When I was sourcing a picture for this article, I found all the AI-related images pretty boring. So I did what I usually do, and riffed a little: “what human thing is analogous to this concept?” is how the mental process goes. And found myself looking for pictures of old-fashioned butlers or valets, in the style of Jeeves, the multi-talented and unbelievably patient manservant who gets his employer Bertie Wooster out of all sorts of trouble.
(I couldn’t find any usable Jeeves and Bertie pictures, but I did find the delightful 18th century painting you see at the top of this post, featuring a very bored-looking dude serving coffee. AI agents will now forever look like that in my mind.)
Is that what agentic AI is, or could be? Yet another tech tool that takes away the friction of being human? For myself, I’d rather plan my own dinner parties, and talk to the people I want to invite, and have the menu flop a bit because I burned the cake. But it might be nice to have an assistant who books meetings in other time zones?
Perhaps I’ll start there and try to build one! I’ll keep you posted.
RESOURCE: A good overview on The Conversation of agentic A: AI agents are here. Here’s what to know about what they can do – and how they can go wrong
Main picture: Portrait of a Venetian Family with a Manservant Serving Coffee by Pietro Longhi, circa 1752. Rijksmuseum, CC0 1.0 Universal.
Other things I have written
Women’s Day ghetto? Count me out – A piece on my much-neglected personal website, The Tidiness Project: how making a special day to honour a class of people is just a form of “othering”.
Why cultural competence is a thing when using AI tools – Generative AI – famously and infamously – gets things wrong. How to deal with that when it comes to representing actual people?
Navigating AI hype: Five newsletters that will help – Everywhere you look, there’s an article about artificial intelligence. Here’s a list of the people I follow to help in navigating AI hype.
How can I help you make order from chaos?
Join the Safe Hands AI Explorer learning circle!
Sign up for my Sensible Woman’s Guide to AI and Content Creation, which comes out fortnightly.
Or sign up for my personal fortnightly email newsletter In Safe Hands (it’s a place to pause and think about things).
Book a free half hour of my time here. Find out how I can help with writing, editing, project management, general clear thinking, technical hand holding, or an introduction to AI.
Contact me via email.