In recent years, “agentic” has become a hot buzzword to describe certain classes of LLM-based systems. It has also become the focus of a huge amount of hype and research. Like so many hype words, agentic is badly abused. The abuse has consequences:
- It sets unrealistic expectations for what these LLM-based systems are capable of – making a bubble burst more likely.
- It muddies the waters of our collective conversation; when we don’t speak intentionally, we can’t think clearly.
- It directs huge amounts of time and money into getting LLMs to do something that they are fundamentally incapable of doing.
- It directs time and money away from the architectures that will in fact deliver the important capabilities being hyped under the “agentic LLMs” flag.
One of the reasons (there are others) that LLM-based “agentic” systems are not agentic is that these systems have no capacity for intrinsic motivation. “Intrinsic motivation” is an internal state of drive, desire, or want that arises from within the system without the need for being explicitly instructed, directed, or prompted by a human user.
Let’s take an example. Ren, my hybrid AI Partner in Care that supports informal caregivers, has a purpose (that I gave it) called “Understand the care circle structure”. What this means is that Ren works to understand the people in a care circle and how they relate to one another. “Elena provides care for her mom, Maria” is the kind of thing Ren works to understand.
When Ren encounters a new person (“Elena”), it checks what it knows about Elena in the context of care circles and finds that it knows nothing. This divergence between Ren’s programmed purpose (“Understand the care circle structure”) and what Ren knows about Elena’s role in a care circle (nothing) triggers Ren’s motivational state (“want to know if Elena acts as support provider”).
I programmed into Ren the mechanisms of motivation. But I do not trigger specific motivational states. I do not prompt Ren, “You are a social system analyst. Your job is to identify the members and relationships between people in a care circle. There is a person named Elena. Execute a multi-step plan to gather information that clarifies Elena’s role in a care circle.”
Ren generates its motivation state autonomously based on its programmed purpose and its understanding of the world.
As I said earlier, “agentic” LLMs have no capacity for intrinsic motivation. In fact, an LLM has no motivational state at all – intrinsic or extrinsic. No goals, no wants, no intentions, no continuity. Even wrapped in “agentic” scaffolding:
- The goal comes from the prompt/system message (exogenous)
- The “agent loop” is external orchestration code
- The LLM itself never “wants” anything between calls
- Shut it down mid-task and restart: it has no unfinished business, no frustrated intentions
A real agent:
- Wakes up with drives/goals already active
- Generates new intentions based on its state and observations
- Maintains “I want X” continuously until satisfied or revised
- The motivational state persists and evolves
So, what are “agentic LLMs” if not agents? The best way to think about agentic LLMs is to think of them as “Smart Macros”:.
- “Macro”: It exists only to execute a (complex) command given to it by an external user. It doesn’t have intrinsic motivation any more than a macro in Excel does.
- “Smart”: It is “operationally autonomous”. Unlike a dumb macro, where every single step is hard-coded by the user, a “smart macro” (the Agentic LLM) dynamically generates the steps (the plan, the sub-goals) in real-time to achieve the final, exogenous goal.
To be clear, I’m not saying LLM-based systems aren’t useful – they’re incredibly valuable. But calling them “agentic” papers over fundamental limitations.
Leave a Reply