Agents as Commodity

AI agents are the AI engineering trend of 2025. It's understandable: For the first time, it's possible to really start to understand the vision of how LLMs can work on tasks for a longer period of time and actually get meaningful work done.

This created a huge boom in new agent offerings across the industry: If we focus solely on coding agents, there are so many different products on the market right now that it's impossible to keep track of. I've been invited to so many coding agents previews, that I can't even remember all the names anymore.

More recently, we see that existing agent offerings start to create SDKs to make it even easier to build on them: Claude Code SDK allows you to build your own agent leveraging some of the context management and tools that make Claude Code so successful. At the same time we see fine-tuned versions of models specifically tailored around agentic tool use.

All of this reminds me a lot of what happened the past year with the model layer. One lab made a breakthrough and many more labs emerged that were able to replicate it. Today, we have more competition than ever in the model layer and it's now perfectly reasonable to have a "best model for a specific task" or even use open-weights models for day-to-day work. Specifically the latter category is in an extreme race towards cheaper and cheaper inference that will eventually converge to only the energy costs.

I think something similar is going to happen with agents very soon. The differences between coding models will become smaller and smaller and more and more solutions will be on the market resulting in the price to eventually converge to only the raw model price.

There are a few more things that need to be figured out still: Security and proper sandboxing, for example, is a big concern. But once we have better primitives in place, it is going to unlock a new abstraction level and I can't wait to play with that.