Kane, super helpful article to cut through the BS and marketing hype. It may be my simple brain having to take a step back, but surely what's important is actually understanding what my business goals are, metrics I'm looking to achieve and why etc etc then figuring out from there what technology/AI I need to implement to achieve those. Are many organisations even ready to get to the level of true Agentic AI/AI agents as per the true definition? Or would other well implemented AI tools actually suffice? Surely it's not just a blanket approach as per some marketing, but actually use case and business value based...
100%. Of course, there’s the argument that many vendors make, which is “as long as useful applications are built, then who cares what it’s called?” I don’t concur with this and actually think that semantics matter when it comes to selling technology.
The right approach is, as you say, to define your requirements and select technologies and approaches accordingly. What many businesses will find is that this approach doesn’t often lead to ‘Agentic’ requirements because much of what businesses do day to day can be described in rules.
So then you have the area of conflict, which is rule based technologies passing themselves off as Agentic to appear cool. The risky zone for me here is that someone buys a system for a non trivial amount of money, believing they have the Agentic future in the palm of their hands, when really, its the same tech as has existed for decades with some generative models handling dialogue management and response generation.
Kane, I really appreciate the effort that you took researching the origins and good definitions for the term. A sad reality is that we won't be able to get rid of the current broad and (deliberately) vague definitions out there, as their purpose is to sell and to some degree it feels like "old wine, in new bottles". So yes, "the term is now just noise".
I feel reminded of the "levels of autonomy" discussion for self-driving cars. Level one being assistance e.g. for lane centering or adaptive cruise control - which does not feel autonomous, but is pretty much a combination of AI for detection and rules to steer or inform.
Interestingly, the "agentic" talk within the Conversational AI space will not be able to live up to the definition based on the property table provided by Franklin and Graesser. Autonomy, temporal continuity (which to me would also relate to sensing, e.g. finding the right time when to proactively engage) and adaptability are not (yet) part of their architecture or feature-set. And while the move from rule-based decision points to more autonomous, agentic "guides" seems to be a logical step forward, it will need to prove its value. Looking forward to your next article!
When Blockchain and Distributed Ledgers were all the hype, a lot of people were attracted by the wild claims and messages of what big change this could create. A lot of them had little experience on how to build successful products or how things actually work in the real world. So the idea, to store everything in this immutable shared, distributed database seemed revolutionary. But it could only handle a fraction of the transaction volume at very high costs. So 10 years in, we have not seen any remarkable innovations other than currency. I do see parallels here. Maybe there is a possibility to create a system that offers some degree of flexibility, autonomy that so far only a human-agent was able to deliver, but the costs for design and running the system, plus the inference-time might mean it will never be more than an experiment. For 99% of the businesses the challenges and the solutions might be in a different space: how can I get 80% automated with 20% of the effort.
I think LLMs in combination with traditional programming could get somewhere close, but as mentioned, that's not strictly the 'true' definition of agency. It reminds me a lot of what Peter Voss talks about, which is that LLMs in their current form are insufficient to reach anything like generalised intelligence, which is perhaps part of what agents need.
They certainly need to be able to respond to novel changes in the environment and learn new skills over time. LLMs don't offer either of these capabilities without supplementing them with other traditional programming capabilities or people.
We've discussed our concern for scaling 'guides' before, perhaps there's an article in that. If you have 300 intents (not uncommon for a large, mature enterprise like yours), then to run these using the 'guides' framework will take 44,850 cross checks to make sure each guide doesn't conflict with any other. That doesn't seem practical to me.
Kane, super helpful article to cut through the BS and marketing hype. It may be my simple brain having to take a step back, but surely what's important is actually understanding what my business goals are, metrics I'm looking to achieve and why etc etc then figuring out from there what technology/AI I need to implement to achieve those. Are many organisations even ready to get to the level of true Agentic AI/AI agents as per the true definition? Or would other well implemented AI tools actually suffice? Surely it's not just a blanket approach as per some marketing, but actually use case and business value based...
100%. Of course, there’s the argument that many vendors make, which is “as long as useful applications are built, then who cares what it’s called?” I don’t concur with this and actually think that semantics matter when it comes to selling technology.
The right approach is, as you say, to define your requirements and select technologies and approaches accordingly. What many businesses will find is that this approach doesn’t often lead to ‘Agentic’ requirements because much of what businesses do day to day can be described in rules.
So then you have the area of conflict, which is rule based technologies passing themselves off as Agentic to appear cool. The risky zone for me here is that someone buys a system for a non trivial amount of money, believing they have the Agentic future in the palm of their hands, when really, its the same tech as has existed for decades with some generative models handling dialogue management and response generation.
Kane, I really appreciate the effort that you took researching the origins and good definitions for the term. A sad reality is that we won't be able to get rid of the current broad and (deliberately) vague definitions out there, as their purpose is to sell and to some degree it feels like "old wine, in new bottles". So yes, "the term is now just noise".
I feel reminded of the "levels of autonomy" discussion for self-driving cars. Level one being assistance e.g. for lane centering or adaptive cruise control - which does not feel autonomous, but is pretty much a combination of AI for detection and rules to steer or inform.
Interestingly, the "agentic" talk within the Conversational AI space will not be able to live up to the definition based on the property table provided by Franklin and Graesser. Autonomy, temporal continuity (which to me would also relate to sensing, e.g. finding the right time when to proactively engage) and adaptability are not (yet) part of their architecture or feature-set. And while the move from rule-based decision points to more autonomous, agentic "guides" seems to be a logical step forward, it will need to prove its value. Looking forward to your next article!
When Blockchain and Distributed Ledgers were all the hype, a lot of people were attracted by the wild claims and messages of what big change this could create. A lot of them had little experience on how to build successful products or how things actually work in the real world. So the idea, to store everything in this immutable shared, distributed database seemed revolutionary. But it could only handle a fraction of the transaction volume at very high costs. So 10 years in, we have not seen any remarkable innovations other than currency. I do see parallels here. Maybe there is a possibility to create a system that offers some degree of flexibility, autonomy that so far only a human-agent was able to deliver, but the costs for design and running the system, plus the inference-time might mean it will never be more than an experiment. For 99% of the businesses the challenges and the solutions might be in a different space: how can I get 80% automated with 20% of the effort.
Very good points made, thank you.
I think LLMs in combination with traditional programming could get somewhere close, but as mentioned, that's not strictly the 'true' definition of agency. It reminds me a lot of what Peter Voss talks about, which is that LLMs in their current form are insufficient to reach anything like generalised intelligence, which is perhaps part of what agents need.
They certainly need to be able to respond to novel changes in the environment and learn new skills over time. LLMs don't offer either of these capabilities without supplementing them with other traditional programming capabilities or people.
We've discussed our concern for scaling 'guides' before, perhaps there's an article in that. If you have 300 intents (not uncommon for a large, mature enterprise like yours), then to run these using the 'guides' framework will take 44,850 cross checks to make sure each guide doesn't conflict with any other. That doesn't seem practical to me.