A new GitHub feature disclosed this week orchestrates AI models to matches developers’ tasks to the best model at runtime, creating what the company calls a more “dependable” coding experience that balances performance and cost considerations.
Experts noted the new GitHub Project HydraFusion is a significant advance because it provides frontier AI intelligence without a developer needing to know or select a model that’s best for the job at hand.
It also adds significant value in the increasingly important realm of AI developer economics. Project HydraFusion “can help with cost efficiency because HydraFusion only calls in extra models when it’s likely to improve the result,” said Jack Rowbotham, senior product marketing manager in Copilot Studio Adoption at Microsoft, in a LinkedIn post.
Before and After Runtime Orchestration
In a typical scenario, developers coordinate models manually: choosing one for a task, asking another to review the work, or escalating a difficult problem to a more capable model.
GitHub previously offered Auto Model Selection features that review tasks and match them to the optimal model for each task, which developers invoke by selecting “Auto.” With Project HydraFusion, that orchestration occurs at runtime, creating a full execution plan, choosing models from multiple providers to draft, critique and revise, or switch to more powerful models to complete a task. A developer chooses HydraFusion once while it manages models and workflow behind the scenes.
HydraFusion evaluates each request and chooses the least complex workflow expected to meet the developer’s needs, using additional model calls only when it anticipates a better result. Developers select HydraFusion like they would choose a model and it, in turn, chooses a workflow that balances quality, cost, and latency for each task.
For each request, HydraFusion chooses one of three execution patterns:
- Single: One selected model solves the task directly. This preserves speed and efficiency when one model can solve the task directly.
- Cascade: A model deemed efficient drafts a solution and a quality gate decides whether to accept that solution or escalate to a stronger model.
- Critique: One model drafts a result then an independent, read-only critic — in the form of a different model — reviews it, and the drafting model revises once. Addition of this independent perspective helps manage tasks where review is more useful than another unaided attempt.
HydraFusion uses reasoning, code generation, debugging, and tool use criteria to select the most efficient execution pattern to meet the quality bar.
When new models become available in GitHub Copilot, they will be evaluated and can be incorporated into its model pool, bringing their strengths to the tasks that they’re best suited to manage.
Project HydraFusion is available to users on all GitHub Copilot plans through /experimental in the GitHub Copilot CLI. Usage is based on the tokens consumed by the models used by HydraFusion.
More GitHub and AI Development Insights:
