Update 18th December: Musing on the novelty of AI agents either pays off by letting me cement thoughts via writing, or I now have raging confirmation bias.
This LinkedIn Post from Professor Dagmar Monett, Director of Computer Science Deparment (Artificial Intelligence, Software Engineering) Berlin School of Economics and Law and University of Berlin is quite obviously frustrated, but masterfully underlined many points here.
This kind of misappropriation of AI terminology to make generative AI appear innovative, as well as any misattribution regarding AI agents to give any relevance to large language models or the like, is so wrong, so annoying, and so disrespectful to a decades-old subfield of AI, that the only thing it does is distort even more all misunderstandings and illiteracy there is with respect to AI! 😡
Please, DO THE READINGS!
Many of the references pre-date my career in any kind of tech governance and I am a not a machine learning veteran, unlike the author. We all see cycles based on our experience. This is an almost 30yr old discussion on what were then called AI agents:
Michael Wooldridge and Nick Jennings' (1995) "Intelligent Agents: Theory and Practice" https://lnkd.in/d4jXqAVR
Michael Wooldridge and Nick Jennings' (1998) "Pitfalls of agent-oriented development" https://lnkd.in/dcW5iBdX from which this quote is a gem:
"[W]e find so-called agents that do nothing to justify the use of the term. [...] Such practices are unhelpful, for the following reasons. First, they will lead to the term 'agent' losing any meaning it has. Second, they rise expectations of software recipients, who will only be disappointed when they ultimately receive a very conventional piece of software. Finally, they lead to cynicism on the part of software developers (who come to believe that the term 'agent' is simply another meaningless management buzzword)."
How prescient! 😭
Just like the path to whatever we call AI has historical peaks, troughs, and switchbacks, the state of my head trying to unpick this hasn't been stable. When 'Briefly Musing' you won't get the same careful references. This is me emptying my head to start the day job.
When I hear the term 'AI Agent' and look at the marketing, I know it was coined to anthropomorphise technology. Maybe it conjures Agent Smith from the Matrix. Maybe just a more sophisticated Alexa. Remembering all the carefully configured voice prompts with very precise wording to make specific things happen (buying groceries, turning lights on, interacting with the Ring camera). In that case an Amazon device, with Amazon software, spliced into Amazon systems, on Amazon infrastructure, with discrete links into third party systems. Reportedly a huge money sink for the company and not the same as running a GenAI enhanced equivalent to drive complex strings of actions in your own network.
Ron Amadeo
Maybe when you hear AI Agent, you go straight to Character AI or other companies hosting creations that mimic well known figures, or people with thematic profiles - Therapist, Philosopher, MMA fighter etc. In the agentic AI stack that is one set of components. Garbage in, garbage out for data is a huge consideration, so is tackling confident errors or persuasively inappropriate responses, along with copyright, privacy, security, and other safety considerations. The bit that interacts with the user, the interface, is the newest part. The spaghetti of tech under the hood... less so.
Maggie Harrison Dupré
To me, because I'm tasked with working out what to consider before banging such things into organisations, it is a mess of effort to work out discrete steps that make stuff happen - a purchase, a customer service workflow, raising an IT ticket. Then enabling the new 'orchestration layer' - the bit that interprets the text, voice, or other generative AI inputs - to trigger those actions. Can inputs be interpreted reliably? Which other systems and processes does it hit? What kind of access is needed? What kind of monitoring? What kind of error handling? What does that mean for downstream systems? What could possibly go wrong?
I explored that in more detail in the last post. There are lessons to learn from rushed implementations, inadequate governance, or unethical tolerance of trade offs. I was exploring the various issues with United Health AI use (rules-based decision engines and predictive models using machine learning rather than the GenAI elements mainly discussed here), long before news about the group CEO broke. I will not be sharing any position on the latter.
Beth Mole
All that is incredibly familiar. I remember so many turns of this handle. Trying to create and configure layers so legacy systems could talk to each other, or (more recently) walking a mile in implementer and developer shoes so I could spot potential gotchas and champion fixes or pivots. Preparing in-house data, systems, and processes to be fit for cloud migration. Enabling new apps to access old databases via intermediary layers that extract, translate, and load data into repositories that the app can use.
Stephen Kaufman
That is what comes to mind when I hear 'Agentic AI'. I hear new tricks and kludges. I hear understated implementation costs and interesting surprises because we are dealing with lots of resource constrained systems and skills shortages. I hear discrete pockets of genuinely beneficial splicing to middle and back end systems, often from companies that have grown up to sell the tools, then I hear more magical language about staff replacement. That hurts my head.
Seeing the wood for the trees is a superpower right now. Knowing enough about trade offs to isolate a valuable use case. Keeping a bead on the macro picture of market state, ongoing volatility, organisational dependence, and all those trade offs for impacted people, processes, and systems.
Caveat Emptor - buyer beware - and Ceteris Paribus - all other things being equal - are doing pretty heavy lifting. By that I mean we are often told about new tech working in AI-centric optimal conditions (or aspirational conditions in some demos), and the most mind-blowing achievements and cautions tend to coincide with funding rounds for vendors.
Frank Landymore
This may sound like bottomless cynicism, but the people tasked with realising tech dreams and using the solutions seem less and less front of mind in much messaging.
It turned out I had more to say than I thought and a few relevant references sprang to mind. Again, what a bright ray of sunshine I am, but the Agentic Emporer is wearing far fewer clothes than claimed, woven at vast expense. Spare a thought for folk helping others to avoid clothing malfunctions when reality lands.
I am ready to be excited by great new developments, but not at expense of some critical thinking. Considered critics seem to be an endangered species as evidenced by the first three pages of any AI related web search, or a glance at LinkedIN.