Ask HN: Who is getting actual value from 'AI Agents'? | Hacker News Reader

Ask HN: Who is getting actual value from 'AI Agents'?

26 pointsNewUser7631213 comments

I saw this post from YC: "From OpenAI’s DeepResearch to xAI’s DeepSearch, we’re seeing the first real push toward autonomous tools that can plan, execute, and complete tasks like research, outreach, and coding with minimal human input."

It got me thinking, are these AI Agents actually useful or valuable to anyone?

I'm fairly technical and I use AI to generate code to make POC projects go faster. I'm not on board with blindly 'vibe coding' though, I think it's dangerous and potentially slower when you get screwed over later trying to fix a large project. I think from other discussions, many on HN are in a similar boat as me - that is, using AI for code generation is certainly valuable, but having an 'agent' just 'work' for you probably isn't.

But anyways, research and outreach, how about those, anyone finding use for these use cases? I've seen a ton of criticism on these deep research products. They seem akin to summarizing the first page of links of a google search, and they're full of sources you'd want to ignore anyways for serious research (e.g. reddit posts). So I'm dubious as to their quality.

Another theme that I think could be useful is using LLMs as advanced/easier RPA tools. Basically, think RPA with some flexibility based on different contexts that can be captured via text. Maybe this is what this 'MCP' hype is all about.

So I'm very curious, who's using what would count as 'AI Agents' (i.e. something more than text chat/prompts), and getting real value from them?

GPT-4o already one-shots Leetcode hard with just a screenshot. I know, it's trained on the answers. But it clicked in my head that if you give the LLM a very detailed specification + tests to the level of detail that Leetcode does, you might actually be able to one-shot any code you like.

Over the weekend, I made a game. It's heavily dialogue based. I wanted a screen and game content that centers around this gameplay. It's not a classic VN, more customized like Ace Attorney.

It's a free game so I need an intro that hooks people in the first 30 seconds. What examples are good for this? Since there's few games like this, I look into movies. I don't want that scene from Inglourious Basterds, that's too fresh. There's a lot of good stuff in classic film, before the 80s that nobody has seen. These movies had to carry themselves on content because they didn't have special effects.

And that's where DeepResearch comes in.

I'm looking for specific films that fit the tutorial of the game that I'm trying to build. Without going into details, there's 4 mechanics I want to squeeze into the tutorial level. I'm looking for specific scenes that capture these mechanics.

One intent is "How do I get players to not skip dialogue like they would with a VN?" AI's interpretation of this is "scripts that encourage impatience". I didn't even know what question to ask, and it helped to rephrase my intent into the right questions then find examples of those.

It wrote a 6000 word essay, plenty of citations, some with clips. I don't have to spend a couple nights watching Psycho or The Night of the Hunter. I can just jump straight to the clips I want. I can skim 24 different movies in the same two nights.

> I think from other discussions, many on HN are in a similar boat as me - that is, using AI for code generation is certainly valuable, but having an 'agent' just 'work' for you probably isn't.

Then isn’t this a terrible place to ask your question? The echo chamber is just going to reinforce what you believe to be true.

Do you think the vibe coders getting a lot done are going to be sitting here arguing with the Never-AI-stans of Hacker News?

(This is NOT a sarcastic comment) Yes, everything else being equal, describing our company as AI Agent for X made it investable and not seem out of place among all the AI Agent for XYZ companies. You don't need to do "agentic" stuff to get value from AI agents. It's honestly a bit gauche to start a company in 2024/25 and not call it an agent.

AI Agents are a lifestyle

The AI companies. Nvidia. That's it. That's the list.

Not right now, but I will as these evolve.

I would love an intelligent agent I can converse with via both voice and text that has access to all my information, privacy be damned.

> as these evolve.

Allow me to introduce you to the first step fallacy. https://thebullshitmachines.com/lesson-16-the-first-step-fal...

I like to use the analogy of building a ladder to get to the moon. The first step fallacy says that building a tall ladder is just the first step to building a ladder tall enough to reach the moon.

I'm aware of this article, and in particular, this lesson. But I don't think it really applies to what I'm talking about.

What I'm talking about isn't attainable via LLMs as far as I can tell, or at least, LLMs alone.

In other words, AI is not just about LLMs.

Fair enough. I'm bearish on AI as a whole, but your point is correct. A true intelligent agent would be handy. Of course, if you're in academia, you'd call it "grad student"

"Agentic" is a fad/hype.

Software applications with more intelligent interfaces get us closer to a bicycle for the mind. LLMs are already a part of some of those.

This is what I’m leaning toward as well, but it’s hard to believe companies are spending billions of compute dollars building agentic tools that no one is using for production projects. Surely there’s some Giganto-Corp, Ltd. out there using this for something other than training their own LLM?

If there were, don't you think you'd have heard about it all over the tech press?

I don't doubt that a lot of it is a fad.

But I think we are in the early stages and we really do not know what is going to work or not. People are hyping their own stuff to get funding or customers.

I am taking things slow. Trying out stuff like PydanticAI and simple mcp stuff.

Reply on news.ycombinator.com