This article covers what changed, what Microsoft measured, what is still missing, and how to switch back if you need to.
What changed: a new harness, not a new model
Microsoft's framing is that a model only reads and writes text. The harness is everything around it: what context the model sees, which tools it can call, what it remembers and what it may touch. A better harness lets the same model finish real work.
Microsoft says the Copilot harness is "battle-tested" at the scale of millions of users, and that GitHub designed it to be fast and predictable. Those are vendor claims. The post also says the harness keeps Copilot going until the task is done and that answers are shorter and more focused. As evidence it quotes one developer, who called it "Less chatter, more quality results for large tasks".
The benchmark: 157 C# tasks
Microsoft's .NET team built a benchmark from real issues in real C# projects. It spans .NET Framework and .NET Core, Azure and Aspire, and WinForms and desktop apps. Both chat experiences ran 157 of its tasks. A task counted as resolved only if its failing tests passed without breaking tests that already passed.
The headline claim is that the new experience beat the old one with every model tested, across capability levels and providers, resolving "up to 21 more tasks". The blog's chart reportedly shows 103 tasks resolved by the old experience and 124 by the new one, using the same model for both. That is 21 more tasks, roughly 13 percentage points of the set.
Some caveats apply:
- This is Microsoft's own benchmark. The post doesn't publish the task list, the test harness or per-provider scores.
- The 21-task figure is a C# result. The post says the engine was also tested on C++, Python and other languages, but gives no counts for them.
- "Up to 21" is a ceiling across models, not an average. Your results will depend on your codebase and prompts.
Microsoft also tells an anecdote. A user reported that long answer choices in Copilot's questions were getting cut off. Asked to help reproduce the bug, the old experience returned manual steps. The new one realized it was running inside the affected chat window and reproduced the bug there. That is a company demonstration, not a controlled test.
New things you can do in chat
Microsoft lists several capabilities that the previous experience lacked:
- Steer mid-task. Type a new instruction and press Enter while Copilot is working. Your message shows as Pending until Copilot picks it up, and you can select the X to withdraw it.
- See results, not just actions. Each step shows what came back. You can expand it, open the results in the editor, or copy the exact command Copilot ran.
- Watch subagents. Delegated work gets its own card showing the model in use and every step, live.
- Work beyond the solution. Point Copilot at any folder from the chat header, or attach whole folders as context.
- Built-in agents. With Agent selected, type
/. Microsoft names/security-review, which runs a security-focused review of your changes, and/rubber-duck, which has a different model critique Copilot's plan before it acts. - Cost visibility. You can see what each request costs. The post's sample screenshot shows 9.0 credits for an eShop basket analysis. That is an illustration, not a typical price.
For IT teams, the credits display and the visible commands are probably the most useful additions. Unmetered AI use is hard to govern, and an agent whose actions you can read is easier to trust.
Still Visual Studio, and still cross-surface
Because the engine is shared, Microsoft says an agent skill written once works unchanged in Visual Studio, VS Code, Copilot CLI and the GitHub Copilot app. Visual Studio-specific tooling remains. Microsoft cites specialized agents for debugging, profiling, build failures, .NET modernization and C++ build speed. It also cites built-in .NET and Azure skills and your existing MCP servers. The post doesn't say which tool versions or configurations each of these needs.
18.11 versus 18.12: reading the version story
The sources don't line up neatly on timing. The blog says the experience is on by default starting with 18.12. Microsoft Learn's Copilot Chat page says that the 18.11 October Update introduced the new chat experience on the Copilot SDK, and that you enable it under Tools → Options → GitHub → Copilot → Copilot Chat, check Use new Copilot chat experience, and restart. The Visual Studio Insiders release notes say that the new experience is now the default, and they list a known-issues section for the 18.12 Insiders release.
The most consistent reading is that 18.11 offered the new chat as opt-in, and 18.12 flips it to the default. No source states that sequence outright, so treat it as inference. If you are on 18.11, you can turn the experience on yourself. If you are on 18.12, check that it is actually enabled.
What's missing, and how to roll back
Microsoft says a few features from the old chat are still being ported, specifically plan mode and chat exports. If you rely on either, or want to go back for any reason:
- Open Tools > Options > GitHub > Copilot > Copilot Chat.
- Clear Use new Copilot chat experience.
- Restart Visual Studio.
Microsoft asks users who switch back to say why. Feedback goes through Help > Send Feedback > Report a Problem, or through Developer Community.
Bottom line
This is an architectural change to Visual Studio's AI assistant, not a new model. Mid-task steering, visible results, subagent cards and cost display should make agent sessions easier to supervise. The 21-task gain is a Microsoft-run C# result and hasn't been independently verified. Teams that depend on plan mode or chat exports should hold off or keep the toggle off for now. Everyone else could try it on a real backlog item and see whether the claimed gains show up in their own code.