Uncharted Waters are difficult to swim in, but it will take time

Uncharted Waters are difficult to swim but it will take time

The fact that our area is open to a diverse spectrum of work is one of the reasons why machine learning has been so successful. I am unable to remember of even a single instance in which a person constructed what they referred to as a machine learning algorithm, and senior members of our community condemned it by stating, "That is not machine learning!" It is true that mathematicians Legendre and Gauss employed linear regression with a least-squares cost function in the early 1800s, which is a significant amount of time before the creation of computers. However, machine learning has adopted these techniques, and we frequently refer to them as "machine learning" in basic courses!

On the other hand, I witnessed statistics departments at a number of colleges looking at breakthroughs in machine learning and saying, "That's not really statistics." This occurred around twenty years ago. This is one of the reasons why machine learning has grown significantly more in the field of computer science than in the field of statistics. Since then, fortunately, the majority of statistics departments have been far more receptive to machine learning in their research.

When I was thinking about how to communicate about agentic systems that leverage design patterns like as reflection, tool usage, planning, and multi-agent cooperation to achieve better outcomes than zero-shot prompting, this contrast came to mind because I was trying to figure out how to talk about them. I had been a part of discussions over the question of whether or not particular systems ought to be considered "agents." I came to the conclusion that it would be more beneficial to conceive of systems as being agent-like to varying degrees rather than having to make a decision in a binary fashion regarding whether or not anything is an agent. In contrast to the word "agent," the adjective "agentic" enables us to consider such systems and incorporate all of them into this expanding trend.

Utilizing design principles that are similar to those used by agents, an increasing number of individuals are developing systems that prompt a huge language model several times. However, there is a gray area between what is obviously not an agent (such as prompting a model once) and what is obviously an agent (for example, an autonomous agent that, when given high-level instructions, plans, employs tools, and carries out several iterative phases of processing).

It is possible to admit that there are varying degrees to which systems can be agentic, rather than engaging in a debate over which types of work should be considered actual agents and which should be excluded. In that case, we will have an easier time including anyone who is interested in working on agentic systems. In addition, we may encourage newbies to begin by constructing straightforward agentic workflows and then gradually increase the complexity of their systems through iterative processes.

Over the course of the last several weeks, I've seen that although those who are not technically oriented and those who are technically oriented both occasionally use the word "agent," the phrase "agenttic" is mostly used by exclusively technical people (for the time being!). Therefore, when I come across an article that discusses "agentic" processes, I am more inclined to read it. This is because the piece is less likely to be marketing fluff and more likely to have been authored by someone who is familiar with the technology.

Let us continue to work on agentic systems and extend a warm welcome to everyone who is interested in joining our working group!