Thinking, Fast and Slow — Daniel Kahneman
We have been building our agent in-house for a couple of months now.
-
The real use case of having subagents working together is that you can maximise lightweight, dumber agents and heavy, smarter agents. Think of the book Thinking, Fast and Slow. Users don’t need the smartest agent all the time because they’re not asking for the hardest questions all the time, so you need a variety of agents.
-
Multiple agents splitting the work amongst themselves is usually “busy work”; these agents have access to the same memory and context and will eventually do the same thing, especially if you have the correct guardrails. So what’s the point of spinning up thousands of subagents when you don’t have much work to break down? Zero value.
-
Routing is more important now than ever — see OpenRouter’s acquisition.
Jev from TypeSafe AI is out now, but we have had small decision models for a long time. I believe the positioning and the virality that followed on Twitter was important in getting people to understand its usefulness.
I explored this when we initially tried building a matching engine internally with Cohere, OpenAI models, etc. They all have their versions of decision models.
My key takeaway from the Jev avalanche is that there is a place for Jev-like models in your stack if a lot of the work you’re doing is some type of classification (not all types of classifications). However, you still need the more powerful models for actual reasoning.
Some teams are already combining the decision models with other powerful reasoning models. I see how that will work.
OpenRouter’s value will become increasingly clear over the next few months as more engineering teams get the education they need and are exposed to different shapes of user problems.
Selah.
Originally published on LinkedIn.