GitHub has launched Project HydraFusion as a research preview for all Copilot plan users within the GitHub Copilot CLI. The tool is not available as open weights or for self-hosting. To activate it, users must run the /update, /experimental, and /model commands to select the HydraFusion (Research Preview) option.
The system treats workflow selection as an optimization problem, analyzing capability signals for debugging, code generation, reasoning, and tool use. It selects the least complex workflow that meets the quality threshold, utilizing additional model calls only when necessary to increase quality.
The orchestration runtime uses three specific patterns: Single, which prioritizes speed; Cascade, which maintains a path to stronger inference; and Critique, which adds an external review process to improve results. The system logs internal data on cost, latency, diagnostics, diagnostics, and role outcomes per leg of the workflow.
GitHub evaluated the HydraFusion policies using Claude Opus 5 and GPT-5.6 Sol as baselines, testing them on three agentic coding benchmarks. One of these is CheckpointBench, an internal multi-turn set based on real Copilot sessions and public commits.
