Decagon uses SGLang in an AutoResearch loop for inference optimization

SGLang's account on X says Decagon runs its inference stack on SGLang and has built an AutoResearch loop for optimizing it.

In that loop, agents diagnose bottlenecks and test candidate changes against replays of production traffic, then build on only the changes that verify as gains.

The report comes from SGLang's own account, the open-source project that Decagon uses, and does not include measured performance results.