<lp>

2025, The Year of Agents?

Published:

Swyx, on Latent Space, mentioned this idea and I knew I had to write about it: “Everyone is saying 2025 is the year of agents, just like they said last year.” It caught my attention, just as it should catch yours. If we’re truly in the year of agents… should we all be looking for applications of agents?

What the hell is an agent anyway?

I’ve seen this word thrown around so much in 2024 that I’ve become somewhat uninterested in it. Everything that everyone is building is agentic - it’s the buzz word to sound credible and cool. I’ve seen it used in so many different contexts that I’ve come to distrust it at face value.

Side note - It’s especially cringe when I hear it from non-technical people 🙂

However… I did stumble across a great tweet by Simon Willison that challenged people to define what an “agent” really is. In this thread, some of the bigger names around the industry weighed in. Of all of them, I prefer Swyx’s answer the most: “agent = llm + memory + planning + tools + while loop.” It’s incredibly simple and easy to comprehend - and I actually agree with it.

So, why does this matter?

As an AI Engineer - it’s my job to stay on top of this. It’s my job to watch the industry and look for applications within my company. I’m especially keen on finding a use-case for agents - but like Swyx and Alesso - I’m not sure we’re there yet. For a company to adopt an agentic workflow - there needs to be a clear benefit over traditional software-defined control flows. I’m not sure that we’re at a place, for most use-cases I’ve seen, where I can trust the LLM to handle the control flow unsupervised.

In fact, I think the thing most people visualize when they hear “agent” is something far superior than current technology allows. Think about the agents you actually have seen… The ones that come to mind the most - and probably the most widely used right now - are coding agents. Think Cursor Composer or Windsurf.

All of them have something in common, that often trumps the automation we imagine: you and I in control. It’s not just a play/pause type of control - there is an entire user experience built around the human in the loop despite every automated action the application takes being an agentic action (break down problem, select tools, loop until done).

Are we imagining the right thing?

I guess that leads us to my final question, as we head into 2025. Agents are very likely to be dominant in 2025 - but does the general population know what an “agent” really is? Do us developers have a common image or visualization of an “agent” - and know how to build around that?

And my spicy take… is “agent” a buzzword that dozens (maybe hundreds) of VC-backed AI companies are using to fabricate the next “frontier” and rake in more cash?

Further Reading

After posting this, I stumbled across a recent HN thread worth reading to see other’s perspectives - https://news.ycombinator.com/item?id=42431361