Gary Marcus Is Right On AI Agents Being Overhyped For Now,
Honey Shows They're Easy To Do

Gary Marcus is a polarizing personality in AI, mainly because he does not agree with the status quo and while he has said things that have been proven wrong, at this moment in time, he isn't wrong about AI agents.

Gary believes by the end of 2025 AI Agents will have just been hype. At this precise second in time I agree with him. Nothing we've seen, not from openai, not from anthropic, manus or anywhere else are making agents for the common man.

If you are a coder ai agents have helped some, let's be real if you're doing something complex claude code is unreliable. I didn't test openai Codex or Google's take, so I won't pass judgement on agents as a category when it comes to code. The creations I've seen agents help with usually are basic games. I've repeatedly said the only thing you can vibe code right now are html sites and small javascript games. When I say vibe code I am using the definition that means, I don't have to know a line of code, I don't have to review the code, I ship it, it works.

Marketing Agents: Limited But Useful

When it comes to marketing, agents are limited in what they can do, but I have to admit they expedite things a lot. When I say they expedite things a lot that is coming from someone who admittedly is not a marketer. I use agents to tell me okay how would I market this. I've found very specific prompts like "I have 25,000 dollars how would I market this with no ads, all organic" and the agent comes up with an actionable list of things I can do, but they do not act on the list for me.

Agents have always been advertised as being able to do the thing for you, and that's where agents are failing, at least with marketing. At this point in time all these agents do when it comes to marketing is review data and come up with an actionable plan. That is something not to be dismissed, but it isn't where we want to be at with AI agents.

What We Actually Want

I want agents and I think you want agents who can learn who you are, and figure out what you want before you even say it and ask if you want them to take action. These would be as Gary says "I want general agents to do my boring tasks", and he is right, they should be able to do that.

The Solution Already Exists: Honey

Here's what's fascinating: effective agents already exist, and they work perfectly. Honey scours the web for coupon codes, grabs them, then when you go to a store that matches those coupon codes it automatically tries to find ones that work for your purchase. That's exactly what people want from AI agents, something that does what they want, instead od telling them how to do what they want.

Since Honey has been around long before the phrase Agentic AI became popular, we know it isn't complicated to automate agents to do things. Perhaps people just like to complicate what things are for views and for influence. Some do it because they don't understand what they're talking about, but more do itfor the views and others try to reframe it to harm their competition.

Google has semi manually been helping people with their flights forever, from scraping their inbox and presenting the data to them in search, if I recall corrrectly. There are plenty of sites to search for deals on flights, so why can't an AI find me the cheapest flight, and get it? When it comes to generalized agents, we should at least be able to generalize to the level that Honey can do things, and Honey doesnt even advertise themselves as agentic AI.

I know that is far more complex than it sounds, but we've shown we can make AI work somewhat for code, so there is nothing stopping us from expanding it into the mundane.

Our Current Reality

Right now, our agentic abilities barely exist. You can tell dreami "look up the book catcher of the rye" then dreami will look at the web and come back and give you it's opinion on what it saw, instead of raw data. It can look up anything, that's just an example. It currently does not provide citations. But that is very limited, it is an AI that can view the web and form its own opinion, which is awesome, but that's hardly the agentic reality we want.

So when Gary Marcus says AI Agents are going to be just hype, and as of July 2025 he is right and guess what? I don't want him to be right, because as Honey shows us, AI agents can be amazing and help you greatly in the most mundane tasks. So, if you are a developer, let's do what we can to make Gary wrong before the end of 2025 and automate those boring tasks, stop focusing on what will make your work faster and focus on what will make people's lives easier and more efficient. Something most developers don't realize is our boring tasks aren't common, in fact most people don't even know they exist.

The Challenge Ahead

Everyone loves saying Gary is wrong, but right now he is on track for being right, and while I don't like it, I accept it. I am not a yes man, if someone says something and they're right, I will say they're right.

If someone has a valid point, it is nonsensical to deny that their point is indeed valid or at a bare minimum not pragmatic to do so.