GitHub’s Copilot weekly releases on September 25, 2026, combine model availability, agent safety, observability, and collaboration updates. Rather than only improving answer quality, the release is about how a team manages a coding agent that runs for longer and touches real project tools.
Local sandboxing in the Copilot app is now in public preview, allowing administrators to limit an agent’s access to files, networks, and credentials. GitHub also added OpenTelemetry configured through enterprise-managed settings so agent activity can flow into existing monitoring tools. One feature defines what an agent may touch; the other helps teams see what it did.
GitHub also added Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 to different Copilot plans. More model choice makes policy more important: cost, task risk, and quality requirements need to be managed together instead of being decided ad hoc in every thread.
The Slack and Microsoft Teams changes focus on context and recovery. Users can switch models inside a conversation and Copilot keeps that choice for the thread. It checks for similar issues before creating a new one and improves implementation-plan status, recovery from interrupted or stale replies, and reconnection after idle periods. Slack also accepts supported files, attachments, and message links as context.
JetBrains adds assisted approvals in public preview: low-risk tool calls can be approved automatically while higher-risk actions prompt the user. Editing an earlier message rewinds both the conversation and file changes before the replacement request is sent. VS Code is gradually adding agent runs in Dev Containers on SSH, Tunnel, and WSL hosts, placing the agent beside the remote project’s actual tools and dependencies.
Together, the changes move GitHub’s coding agent from a chat surface toward a managed execution environment with permission boundaries, telemetry, state recovery, and a pause before higher-risk actions. For enterprise adoption, that is closer to the operational problem than a single benchmark score.
Most of the capabilities are previews or rolling out gradually, and availability depends on plan, administrator settings, and workspace. Before using them in production development, teams should test how the sandbox affects builds, tests, package registries, and deployment tools, then document repository, branch, review, cost, and cancellation rules.
