These days, I came across a project on HackerNews proposing to wrap agent architectures with a deterministic control layer, composed of sets of formal rules—however many may be necessary—to mitigate issues such as hallucination, infinite loops, or goal drift in these components.
Without debating the merits of the idea itself, I liked a provocative comment in the thread that questioned the intention of going back to writing formal rules and instructions for a working domain, when the goal of adopting agents was precisely to avoid having to write them.
As an ordinary human, I always latch onto a pattern when I see one. The current buzz around projects applying LLM-based agents to commercial activities strongly reminds me of the attempts to use blockchain for a wide range of problems in the early 2020s.
Blockchain is a technically fascinating concept and, together with the other parts of the Bitcoin protocol, helped make “Byzantine consensus” viable. It solved a major (and valuable) specific problem. But its use ended up being limited to that purpose.
In 2020, I was leading Getmore.com.br and I remember having many discussions—some intense—about how it made no sense to use blockchain in building incentive and loyalty programs, as well as in many other domains beyond Bitcoin. History, in retrospect, proves me right, but not riding the hype cost us business and opportunities at the time. Sometimes you have to play the fools’ game.
I believe that the strength (and the limit) of using agents will be established in solving elementary problems of assistance in language production, including software development, the production of general multimedia content, as well as domain-specific content—for example, medical or legal content.
I agree that this represents a broader spectrum of problems than the previous blockchain context. The limits, in this case, would not be of field, but of scope and autonomy. In the end, we will probably have to continue formalizing and describing the domain.
Without debating the merits of the idea itself, I liked a provocative comment in the thread that questioned the intention of going back to writing formal rules and instructions for a working domain, when the goal of adopting agents was precisely to avoid having to write them.
As an ordinary human, I always latch onto a pattern when I see one. The current buzz around projects applying LLM-based agents to commercial activities strongly reminds me of the attempts to use blockchain for a wide range of problems in the early 2020s.
Blockchain is a technically fascinating concept and, together with the other parts of the Bitcoin protocol, helped make “Byzantine consensus” viable. It solved a major (and valuable) specific problem. But its use ended up being limited to that purpose.
In 2020, I was leading Getmore.com.br and I remember having many discussions—some intense—about how it made no sense to use blockchain in building incentive and loyalty programs, as well as in many other domains beyond Bitcoin. History, in retrospect, proves me right, but not riding the hype cost us business and opportunities at the time. Sometimes you have to play the fools’ game.
I believe that the strength (and the limit) of using agents will be established in solving elementary problems of assistance in language production, including software development, the production of general multimedia content, as well as domain-specific content—for example, medical or legal content.
I agree that this represents a broader spectrum of problems than the previous blockchain context. The limits, in this case, would not be of field, but of scope and autonomy. In the end, we will probably have to continue formalizing and describing the domain.