Agentic AI

I'm Building An AI Agent. Except I'm Not. And Neither Are Most Of Them.

On July 4, 2026, I successfully deployed the alpha version of B.E.N.N.Y., the AI agent that I am building on another project I am currently working on, J.A.L.E. (Job Automation & Logistics Engine), which I have mentioned in a previous article.

However, B.E.N.N.Y., like most other AI agents, is built on if-else logic on top of an LLM layer. By admitting this, I am forced to admit B.E.N.N.Y. isn't really an AI agent, at least not yet.

At the time of writing this article, B.E.N.N.Y. has so far successfully performed two of the following tasks:

1. Handle interactions with a potential client and gather details, such as their business information, the service they're requesting, a brief description of what they want, etc.
2. After B.E.N.N.Y. gathers all the information, it will ask the user to confirm the information. Once it's confirmed, the lead is then submitted via email, where either one of my team members or myself will review it to ensure its authenticity, validity, and whether it fulfills other requirements, such as legality, our business goals, etc.

In the future, B.E.N.N.Y. will also serve as a monitor where if there are any irregularities in timesheets, such as abnormal shift hours, the amount of time it took to earn a specific number of hours, etc., it will flag it for human review so that an administrator or HR rep could review it before bringing in the employee in question.

There is a common term used to describe AI agents who have pre-determined workflows when they’re deployed to production for clients to use. It’s called “agent washing”.

Gartner, a research and advisory firm, first coined the term back in 2025, utilizing it to describe the rebranding of AI assistants, automation tools, and chatbots that have already been built sometime prior before the AI agent landscape changed. Gartner has stated that only 130 of the thousands of agentic AI vendors that claim to be selling authentic variants exist.

What I described earlier about Benny also falls into the same category. The workflow is pre-defined, and so is the logic, as it is conditional and will execute specific steps depending on the information inputted by the client. The latter is handled by the LLM layer, which will interpret the information before deciding to qualify their request as a lead. In other words, the only entity responsible for making the decisions is the if-else structure that has been set up to handle the different scenarios.

Because these misclassifications are usually a result of pressure on companies who sell these products, in addition to the demand of clients wanting to use them in their daily business operations, this has led to increased scrutiny not just from the aforementioned parties (buyers and investors), but also from regulators.

According to a post written and published by Chary Chandrasekhar and colleagues for the Harvard Law School Forum on Corporate Governance on April 16, 2026, the overuse of the term "AI agents" and other similar terminology creates an inherent legal risk for those using it for marketing purposes.

This begs the question: if the majority of most "AI agents" are just a pre-defined set of configured steps and instructions, then what truly qualifies a product to be an AI agent?

For it to truly qualify as one, they need to be able to dynamically determine what information it needs. Furthermore, it needs both the capability and the autonomy to determine which tools or decision processes to invoke, and when chatting with a user about a specific request they have in mind, adjust its approach accordingly as it sees fit. That's my goal for B.E.N.N.Y.

This perspective isn't new. In a 1994 study by MIT researcher Pattie Maes, she argued that a true AI agent does not wait for its next set of instructions, but rather initiates, monitors, and acts based on the patterns it has learned over time.

With that process in place, the AI agent has the autonomy to decide what the best course of action is, based both on the request it receives from the user and the patterns it has learned from previous interactions. However, to ensure the safety of both the consumer and the producer of the product, the autonomy should be bound by human-in-the-loop practices. This approach allows the people in charge of monitoring the AI agent's performance, review its conversations, and decide whether it is truly safe to continue utilizing and distributing their product.

Further evidence supporting this point comes from an April 2026 paper written by researchers Kwon Ko and Hyoungwook Jin, who are from Stanford University and the University of Michigan, respectively. They cited a 1996 article written by Michael Wooldridge and Nicholas Jennings stating that AI agents (called intelligent agents at the time) were defined through four categories, which were:

1. Autonomy, where agents operate without the need for human intervention, making decisions on their own to a certain extent.
2. Reactivity, which allows agents to perceive their environment and responding to the changes in a timely manner.
3. Pro-activeness, which directly connects to the reaction the agent experiences, allowing them to take the initiative without the need for explicit instructions.
4. Social ability, allowing agents to interact not just with human users, but other agents as well through a shared communication protocol.

By using these four categories, I will argue that most products currently marketed as AI agents only satisfy one or two of them, or in some cases, 3 at most. Because they only partially meet the criteria, they often fail to qualify as true agents.

Ko and Jin demonstrate how important this criterion is when it's distributed to products using agentic systems. For example, they stated that a lamp that adjusts its brightness when someone is tired satisfies reactivity, whereas a bed that detects disrupted sleep patterns and signals a potential health concern satisfies autonomy. When we look at today's digital landscape, and the products that are currently being marketed as AI agents, we realize that most of them don't meet the criteria. In other words, they fall significantly short of the definition.

So where does B.E.N.N.Y. stand when applying the criteria check?

After deploying the alpha version on J.A.L.E. (which will be covered in another article), I put it through various tests to see if it was able to fulfill the four criteria. However, I want to state that in its current state, B.E.N.N.Y. is only capable of one task, which is to receive and qualify leads. The timesheet monitor feature is currently in development and has not yet been deployed.

When I ran the tests, I realized that B.E.N.N.Y. only met two of the four criteria that Ko and Jin have cited in their paper:

1. Social ability: B.E.N.N.Y. passed this one as it was able to interact with me in a natural manner and has gathered information from me, such as my name, email address, city and state, etc.

2. Reactivity: B.E.N.N.Y. was able to respond to and adjust its responses accordingly depending on what the user has typed into and submitted from the input box. However, since it only reacts when a user initiates, it partially qualifies for this check.

As far as autonomy and pro-activeness are concerned, B.E.N.N.Y. fails to qualify for both. At the time of writing, B.E.N.N.Y.'s decision structure is pre-determined through the if-else statements I wrote into its system and is not able to flag, initiate, or pursue any conversations or goals as it has not been set up with those capabilities yet.

Ultimately, I want to admit that as I wrote this article, I finally realized that for a product to truly qualify as an agent or any similar variant, it must meet specific criteria as they should be able to make their own decisions and carry out their own tasks without always needing human intervention while simultaneously keeping a human-in-the-loop architecture.

We can't simply market products just because they're able to interact with users on a conversational level and call them AI agents. If they're not able to dynamically adjust their approach, then they don't qualify as one. Automated chat responses, the kind you get when contacting customer support services to ask for more details about a specific product or making a payment to a bank via phone all come with pre-defined responses, most of them anyhow.

When a business buys a product that is supposedly an AI agent, they expect it to do all the work for them, when in reality, it fails to meet expectations. The reason is simple: they were hyped and oversold when they were never set up with the autonomy and the agency to allow them to carry out their own decisions and dynamically adjust its approach. As a consequence of this marketing, Gartner is reporting that 40% of agentic AI projects are expected to be cancelled by the end of 2027.

As someone building an agent myself, it would be easy for me to call B.E.N.N.Y. one. However, after doing the research needed for this piece, I now realize that I have serious work to do to ensure that it meets all of the criteria. I want to make my products, not just B.E.N.N.Y., not just marketable, but honest in what it actually does, and not turn it into something as meaningless as a product supposedly running on blockchain.

Correction on July 9, 2026 at 6:43 AM: Fixed several typos, including a misspelling of Nicholas Jennings's first name. The article has been updated to reflect these changes.