Why Nobody Has Solved Continual Learning
Continual learning from agent traces remains fundamentally unsolved, with SFT and RL each having critical failure modes.
“no matter what anyone tells you, like, the train box is like not solved at all. It's really, really, really hard, and everything that you do has problems and trade-offs.”
A LangChain researcher explains why the 'train' step in the agent feedback loop is unsolved: SFT is dangerously off-policy (models learn from frontier traces but fail to recover from their own errors), while RL is bounded by the model's existing sampling distribution. The practical implication is that open-source labs likely use distillation to warm-start models before RL, a pattern the speaker claims to have replicated internally. This matters because it sets realistic expectations for anyone building self-improving agent pipelines.