Skip to content
Theme
OnDevice Tools

AIAI Tools & Apps

VS Code makes Copilot harness the default for new agent sessions

The VS Code Blog says new agent sessions in VS Code now use the Copilot harness by default, while the older Local harness enters a three-month transition.

VS Code illustration showing the Copilot harness selected for new agent sessions

Key takeaways

  • New agent sessions in VS Code now use the Copilot harness by default.
  • The older Local harness is deprecated, with a three-month transition before planned removal in February 2027.
  • One online comparison reported 14.2% lower token use per active device and 9.6% lower estimated processing cost.
  • Offline evaluations found fewer errors, fewer stalls or crashes, and more completed tasks.
In this article
  1. What changed
  2. Why the shared runtime matters
  3. Efficiency and reliability results
  4. Benchmarks and new capabilities
  5. Rollout and feedback

What changed

According to the VS Code Blog, new agent sessions in VS Code now use the Copilot harness by default. The change is presented as a move from the older Local harness to a shared runtime, the common software layer behind Copilot products.

The Local harness is deprecated. It will stay available for three months while users move across, with removal planned for February 2027 unless an organisation restricts its use. In local projects, the Copilot harness continues to run on the developer's machine within VS Code.

Why the shared runtime matters

Illustration: code, Copilot logo, AI, settings and the labels Copilot Cli, Copilot Harness, Copilot SDK

A harness gives the model context and tools, manages steps such as code changes and test runs, and sends outcomes back. VS Code says the Local harness came first, created for interactive coding before a shared Copilot harness existed. That earlier work influenced review of agent changes and the use of tools from VS Code extensions and MCP servers, which provide tools to agents.

The Copilot harness relies on the Copilot SDK, a development kit, to bring together editor integrations supported by VS Code and that shared runtime. That runtime also appears in Copilot CLI, the GitHub Copilot app and other Microsoft products. It is included in VS Code and supports current Copilot plans. When that shared runtime is fixed or improved, VS Code and other Copilot clients can gain from the work. This can free engineering time for features, fixes and workflow improvements tied to VS Code.

Efficiency and reliability results

VS Code reports that one online comparison during the Stable rollout found lower resource use with the Copilot harness as the default. In that analysis, Copilot Business and Enterprise users on VS Code Stable in October were split into groups, with new chats using either the Copilot harness or the Local harness.

Token use, measured in the units models use to process text, was 14.2% lower per active device in the Copilot harness group. Estimated processing cost was 9.6% lower, and the reduction was described as statistically significant. This estimate covers the AI work needed to handle agent requests. It is based on token usage, cached input, which is reused context, and rates for each model, rather than billing records.

Offline evaluations also favoured the Copilot harness. Across repeated runs, the team recorded 12,845 side-by-side comparisons. Each comparison used the same task, model and reasoning effort. The set included 112 long coding tasks from the DeepSWE benchmark, covering seven models. The Copilot harness recorded fewer errors, fewer stalls or crashes, and more completed tasks. The tasks were multi-part feature requests in open-source projects, including an example involving standards-based deprecation headers for FastAPI.

The company distinguishes online evaluations, which reflect normal work in the editor, from offline evaluations, which repeat realistic development tasks under controlled conditions. The offline work covers comparisons specific to VS Code and benchmarks for the shared runtime.

Benchmarks and new capabilities

The blog also points to published benchmarks for the shared runtime. Those tests set Copilot CLI against harnesses supplied by model providers, with matching models and tasks. VS Code says the tests found similar task completion rates, with fewer tokens needed in most setups. Copilot CLI uses the same runtime that VS Code brings in through the Copilot SDK.

The shared foundation is said to enable features beyond the Local harness. These include Copilot Fusion for work that uses multiple models, enterprise permission and sandbox policies enforced by the runtime, and Agent Host sessions that can stay active after the project window closes as long as VS Code remains running.

Rollout and feedback

The Copilot harness was expanded gradually. VS Code says the rollout started internally, then reached Copilot plans, VS Code Insiders and Stable, plus editor and Agents-window workflows. At each stage, metrics and feedback were reviewed before further changes were made.

Some improvements have already shipped. VS Code version 1.143 cut delays linked to MCP servers that did not complete tool loading. Version 1.145 made it easier to reopen and browse long chats, and it shows terminal output as it becomes visible. The blog says sampled first-request results still indicate that initial waiting times can be reduced, and work on that delay continues.

For quality checks, the team looked at measures including completed turns, errors, cancellations and delay. It also split first requests from later requests, which showed different areas for improvement.

Sources

  1. The Copilot harness: Building on a shared foundationVS Code Blog

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.