GitHub HydraFusion: Orchestrating Models

Auto model selection answered the question, “Which model will handle this task best?”

The question was static: the choice was made once, before work began.

Real development tasks do not work this way: code drafting, logic review, and final editing require different strengths, and no single model excels at all three. GitHub formalized this as an architecture: an execution plan, model roles within the plan, and a cascade point.

For a company that has already adopted Copilot or a similar assistant, this means buying the “best model” is no longer the solution—the solution is a pipeline with switching logic.

The mechanics are simple on paper and difficult in production.

A fast, inexpensive model writes the first draft.

The second model—not necessarily the same as the first—acts as a critic, finding logical gaps, style violations, and overlooked edge cases.

If the critique shows that the task is more complex than expected, the system cascades to a more expensive frontier model instead of running it for every request.

The economic impact is measured in TTU: the time from a developer’s request to a working result decreases because the expensive model is activated only where it is genuinely needed.

This is a case where the simple outcome is “code ready quickly and without gaps”

- requires complex routing engineering under the hood.

Orchestrating multiple models simultaneously creates a second task: managing this fleet. On the same day MAI-Code-1-Flash was released, GitHub announced its deprecation in all Copilot modes: chat, inline edits, autocomplete, and agent mode. Teams that hardcoded a specific model into a pipeline or policy face an unplanned migration without time to prepare.

The response to this class of problems is enterprise managed settings for Copilot in JetBrains: administrators gained centralized control over plugins, MCP server access, telemetry, and permission modes.

This indicates that the more models and integrations involved in development, the faster a company without a single control point loses visibility into what is happening.

Meanwhile, CodeQL 2.27.0 added native Linux ARM64 support and a new security query for Rust, expanding coverage for Java/Kotlin and C#.

The growing number of platforms and languages for which code is generated—including by models—requires static analysis to evolve with them rather than lagging by a quarter.