Decagon AutoResearch boosts inference serving throughput 2.50x

Decagon engineers describe AutoResearch, an automated agentic system that continuously diagnoses and optimizes inference serving stacks and speculative decoding drafters. By running high-fidelity workload replays against production traffic patterns and coordinating subagents across parallel experiments, the system achieves throughput gains up to 2.50x across various model endpoints. Open article →