GitHub HydraFusion Orchestrates Multiple AI Models for Coding Tasks

GitHub has launched Project HydraFusion, a research preview designed to dynamically orchestrate multiple AI models for coding tasks within the GitHub Copilot CLI. This new system moves beyond the static selection of a single AI model per request, instead constructing an execution plan tailored to each specific coding prompt. HydraFusion can, for instance, utilize one model for initial drafting, then employ a second model to critique that draft, or escalate to a more powerful model if the initial attempt fails to meet a defined quality threshold. The models integrated into HydraFusion can originate from various providers, offering developers flexibility in their AI toolchain.

HydraFusion is currently accessible as a research preview to all users with GitHub Copilot plans, but its deployment is limited to the GitHub Copilot CLI. There is no provision for open-weight models or self-hosted deployments at this time. To enable HydraFusion, users must execute the commands `/update`, then `/experimental on`, followed by `/model` and selecting "HydraFusion (Research Preview)". The billing structure for HydraFusion is based on the total tokens consumed by all models invoked within a particular workflow, with each model being charged at its standard rate.

The core functionality of HydraFusion is built upon "Auto model selection," a feature previously introduced by GitHub in early 2026 to intelligently match coding tasks with the most suitable single model. HydraFusion elevates this by treating workflow selection as an optimization problem. It analyzes capability signals related to reasoning, code generation, debugging, and tool utilization. Subsequently, it identifies the least complex workflow that is predicted to achieve the desired quality standard, reserving additional model calls only for instances where they are likely to yield significant improvements.

For each incoming request, HydraFusion currently employs one of three distinct execution patterns. The "Single" pattern involves a single selected model directly addressing the task. The "Cascade" pattern begins with an efficient model drafting a solution, which is then subjected to a quality gate; if the draft is not accepted, the task is escalated to a stronger model. The "Critique" pattern involves one model drafting a solution, followed by an independent, read-only review from a model belonging to a different family. The drafting model then revises its output once based on this critique, mirroring the review process seen in the "Rubber Duck" pattern. Each of these patterns offers a different trade-off between the quality of the output and the associated computational cost, with "Single" prioritizing speed, "Cascade" maintaining a pathway to more powerful inference, and "Critique" introducing an external perspective where review is deemed beneficial.