Don't Build AI Agents -- Do This Instead

Major AI Agents Groupthink BS is going on.
There's a better way to practically leverage LLM's.

AI agents is a cool idea that grabs our imagination. It gives the AI system an even more human like feel to them and it fits within our previous mental models.

AI agents definitely have practical use cases. However, when working to solve real world problems they aren't nearly as practical as what we're being sold by thought leaders in the AI space.

Certainly, not all LLM value will converge on AI agents as many top leaders are predicting.

Anthropic's "Building Better Agents" article:

Major VC firm saying 2024 was the year of agents and 2025 will be the year of agent swarms:

Is Strawberry/Q*/o1 AGI? What everyone's missing:

GPT-X Use Case Fallacy:

#aiengineering

[RaviSharma-vutb]

Breaking things down into smaller chunks is a great advice. With LLMs being able to do a lot of things, how you frame and break down problems becomes very important.

[_meme_]

Ok focus on code not on agents. Agree. But what in 2 years when agents take care of all the code? All these models getting better every day.. We need to prepare to orchestrate complex pipelines, not only with software but with full systems. We need some really good strategic planning not only to keep our jobs but ourselves as species

[Gismotronics]

I fully agree about the 'pick problems into smaller pieces'. I recently started tinkering with AI coding assistants built into VS Code. Giving the AI coding assistant too bigger problem is inviting a downward spiral from which there will be no recovery - except to restore from the last git baseline.

[jonasmischke]

Can you please get more into detail what you mean at 14:30. Whats the better alternative to LLM-Centric?

[gcg]

An interesting take, if I understand correctly using a deterministic framework to steer/guide/contain an undetermined LLM AI? Have I understood you correctly?

[michaelcbond]

As an AI enthusiast, nut non coder. I want to work with the right components to optimize my efforts and my client's efforts.

I was thinking the multi agent systems get clunky. How do you refund us think about building process vs an agent vs a workflow?

[knecting]

15:43 he finally begins to answer the question.

[nexusphreez]

This is essentially what I've done. I've created a program that allows me to create nodes that do specific things when engaging with LLMs or just processing the output or input before sending it to an LLM or receiving from an LLM. I can do things like web scraping, right long form content, books that kind of thing.

[haroldpierre]

I’m not a coder and haven’t written code in over 30 years. I’ve been hearing a lot about AI agents and want to get involved, but I lose interest when tutorials show endless lines of code and dependencies. I rely on AI to help me generate code because I don’t have the technical knowledge. While it’s been frustrating at times, I’ve still managed to create some useful solutions. What we need is a straightforward, easy-to-use AI agent solution. If that’s available, I, along with many others, would use AI agents like crazy.

[RedCloudServices]

I think Enterprise software is about to shrink, if not, disappear completely in the next decade. Agreed - I notice how the Top vendors are in a panic shoving “AI” into their Enterprise stacks. But can DIY LLM agents, tools, functions, etc replace software stacks? 🤷‍♂️

[VectorGraphicStore]

I would agree with you about Sam Altman's statements, if anything, he's hurting OpenAI by creating unrealistic expectations from the people who are taking what he is saying as the real deal. He's exaggerating, big time. It's like dude, your models still lie and make up stories, they don't listen all that well, but they are good at predicting what words should come next based on context. Yea, Altman is literally doing what he is doing so that the money keeps flowing in from investor's. Because if any of the investor's used it in any real-world scenario. They would absolutely see just how limited it really is. I'm not naive enough to say that OpenAI doesn't have some super advanced models hidden away. But whatever they are feeding the consumer is only a step above content spinners. Whatever infrastructure they are building on top of these models may be designed to limit their consumer models. I've been using it and I can assure you that it's inconsistent, lazy, being intentionally limited, and I've yet to build something with it that really actually works. Due to the fact that it still needs you to implement your own guardrails.

[GpWUrbs]

Some great points. A tool video would be useful yes. If you read through langgraph docs I feel like tools and agents are almost always the go to, even though I wasn't sure if thats always the best architectural choice.

[EduardsRuzga]

I am kinda staying away from agents due to similar views as yours. LLMs are not ready for fully agentic work and is just marketing hype that pitching one of two lies. It aether does not work or is not really an agent.

My subscribe for making such video.

[blarvinius]

Maybe in 2025 someone will create something that no one can deny is an ACTUAL agent. Then all this wishful / marketing word creation can be put behind us. ❤

[patruff]

Yes! The low/no code thing has been bugging me so much. People act like they can do so much and they can but it's basically giving low-level access to do low-value work. Anyone want to buy a chatbot?!

[michael_gaio]

I think you are washing over the important distinction between ai agents and ai systems that do not have agency. that distinction is not superficial

[michael_gaio]

agent means = has *agency* …can make decisions and actions on it’s own. it’s that simple

[aiamfree]

lol the word orchestrate is hilarious… it’s just while loops with switch statements and if checks 😅

[context_eidolon_music]

Just don't build anything. Your product will always be obsolete in six weeks or less.

[adanpalma]

To me I have not seen a real business process on agentics on charge. No one I mean a serious and profesional solution. Alwais is summary videos, make some appoiments, or search, develop and supervise but are trivial problems.