GitHub just gave Copilot a new multi-model orchestration layer called HydraFusion. Instead of putting every coding task through a single AI, HydraFusion first figures out what needs to happen — then decides which model to call.
Three execution modes are currently available. In Single mode, a straightforward task goes to one model to handle end to end. In Cascade mode, a cheaper model takes the first swing, then hands off to a stronger model if the results don't pass the bar. In Critique mode, one model does the initial work while another model from a different family plays the reviewer, spots problems, and sends them back to the first model for revisions.
GitHub benchmarked HydraFusion against Claude Opus 5 across three coding-agent tests. HydraFusion came out 4.9 percentage points higher on TerminalBench 2.1, dipped 1.5 percentage points lower on DeepSWE, and landed just 0.1 percentage point behind on CheckpointBench — effectively a tie. The cost savings are starker: 67%, 36%, and 65% lower, respectively.
The design philosophy echoes Sakana AI's Fugu project. Both approaches turn "which model should I pick?" into "how do I coordinate several models?" The key difference is that Fugu is itself a trained orchestrator model that learns how to call other agents and can even invoke itself recursively. HydraFusion, by contrast, sticks to three fixed routing patterns — Single, Cascade, and Critique — which reads like multi-model scheduling baked directly into Copilot.
HydraFusion is open as a research preview for all GitHub Copilot plans. You'll find it under Copilot CLI's experimental features. GitHub says it's currently best suited for single-turn coding tasks that can be stated clearly; longer, multi-turn workflows are still being refined.
Source: GitHub Blog
Dongcha Beating: Telegram channel · Discussion group