GitHub published its weekly Copilot changelog on September 10, 2026, headlined by Project HydraFusion, a new experimental feature in Copilot CLI that performs what GitHub calls “automated semantic routing between local, cloud, and compound models.” Rather than a developer manually picking a specific model for a coding task, HydraFusion is selected like any other model option and then chooses, per task, a workflow that balances performance, cost, and latency on its own. GitHub placed the feature in Copilot CLI’s experimental tier, meaning it is opt-in and not yet a default behavior.
The same release added a Jira integration to the Copilot app: teams can bring Jira issues into a shared canvas, decide which ones move forward, and let Copilot carry that context through investigation, implementation, and pull-request preparation - collapsing a ticket-to-PR workflow that previously required manual handoff between Jira and the IDE. In VS Code 1.137, GitHub introduced public-preview support for recurring agent task scheduling (hourly, daily, weekly, or on-demand), an experimental voice mode for real-time spoken interaction with Copilot, and an Agents window that can now show issue and pull-request detail without an open repository. JetBrains users got expanded enterprise controls, letting administrators centrally set sandbox policy - enablement, filesystem and network access, proxy settings, and macOS Keychain access - across an organization.
Taken together, the release is less about a single flashy capability and more about operational maturity: scheduled recurring agents, centrally governed sandboxing, and automatic model routing are the kind of controls an enterprise rollout needs once agentic coding moves from individual developer experimentation to a fleet of agents running unattended inside a company’s infrastructure. HydraFusion in particular is GitHub’s answer to a problem every multi-model coding tool now faces - developers should not have to know which model is cheapest or fastest for a given task - and it lands the same week Cursor shipped its own coordinator-and-subagents model in Projects, underscoring that automatic task-to-model routing is becoming a contested feature across coding-agent vendors rather than a one-off idea.
The caveat is that HydraFusion ships experimental and CLI-only, with no published benchmark showing how its automatic routing compares to a developer manually choosing the right model - GitHub’s post describes the mechanism and intent but not measured cost or latency savings. It is a capability to watch rather than a proven efficiency win yet.