From training LLMs to training agents
The centre of gravity is shifting. Here is what actually changes when it does — and why the runtime comes before the training method.
Foundation model training was the biggest problem in machine learning for five years. It is still a serious problem. It is no longer the biggest one.
The biggest problem today is what happens when the model is only one component of an agent — a caller that has to plan, use tools, recover, stay safe, and leave an audit trail. Training that agent is a different problem from training the model inside it.
The runtime comes first
You cannot train what you cannot measure. You cannot measure what you cannot replay. You cannot replay what you did not version. That is the order: runtime, measurement, training method. Not the other way around.
We are honest that trace-based learning — using successful traces to train next-generation agents — is an open problem at Xplore Lab. What we ship is the runtime that makes it possible to collect and verify those traces in the first place.