Visual Studio Code 1.147 shifts the editor further toward managed, persistent agent work. Microsoft’s release notes introduce scheduled and saved requests, decision-model routing, visual responses, browser control, remote previews, tighter enterprise policies, Markdown editing changes, platform deprecations, and a substantial set of reliability fixes.
For CTOs and engineering leaders, the signal is clear: agent-first development now requires the same operational discipline as any other execution environment. Scheduling, browser access, model selection, sandbox permissions, and alternative harnesses all create controls that platform teams must test before broad deployment.
1. Agent Sessions Add Scheduling, Drafts, and Persistence
What it is
VS Code 1.147 introduces experimental Send Later support in the Agents window through sessions.sendLater.enabled. According to Microsoft’s release notes, users can select a date and time from the send-button menu or ask the agent in plain language to schedule a one-time, text-only request.
Scheduled work behaves differently depending on its origin:
- New session: The request waits in the Drafts section.
- Existing chat: The request appears in that chat’s Scheduled area.
- Busy agent: The request waits its turn rather than interrupting current work.
- Unavailable host: A missed request runs after the local agent host becomes available again.
The feature works only with This Computer sessions and requires chat.automations.enabled. Microsoft explicitly states that scheduling does not wake a sleeping computer.
Draft management is a separate experimental feature controlled by sessions.drafts.enabled. VS Code now preserves multiple unsent requests, including their text, attachments, inline references, workspace, agent, model, and other configuration. Drafts survive reloads and restarts, and a sent draft remains stored until the request is accepted.
Turning off sessions.drafts.enabled hides the Drafts section without deleting saved drafts. This is persistence, not merely a larger input history.
Microsoft also added GitHub Environment sharing for connected SSH hosts, WSL distributions, and Dev Tunnels. A user can select Share as GitHub Environment in the Remote Environments editor, allowing supported Copilot applications or another device using the same GitHub account to pick up the remote environment’s agent sessions.
Dev Container users gain experimental session persistence through dev.containers.stateVolume.enabled and chat.agentHost.devContainer.stateVolume.enabled. Microsoft says these settings store user and agent state on a Docker volume so chat sessions survive container rebuilds. The volume is assigned when the container is created, so an existing container must be rebuilt once after the settings are enabled.
Inbox mode also gains two triage controls. Users can acknowledge a session without marking it read, finished, archived, or closed; new activity returns it to the Inbox. Workspace and environment pickers now filter every section of the session list, and the selected filter persists across reloads.
Why it matters
Send Later is useful for checking a completed build or resuming a review, but it is not a hosted scheduler. Your team must account for machine sleep, host availability, local automation settings, and the fact that scheduled work can execute after its intended time.
Draft persistence reduces context reconstruction when developers switch between tasks. Platform teams should still test how saved attachments, workspace references, and model selections behave under their data-retention requirements.
Persistent Dev Container state changes container lifecycle assumptions. Teams that previously treated every rebuild as a clean removal of local agent history should review Docker volume retention and cleanup procedures before enabling the feature broadly.
2. Model Routers and Comparisons Expand Model Choice
What it is
The experimental chat.copilot.modelRouter.enabled setting enables decision model-based routers in Copilot harness sessions. A router is a YAML file stored under ~/.agents/model-routers/ and appears in the model picker.
Microsoft says the decision model gpt-6-luna-decisions matches each request to one of the routes defined by the user. A route can, for example, direct multi-step implementation work to opus and quick questions to gpt-6-luna.
Users can invoke /model-router in a Copilot harness chat to describe the desired routing behavior and generate a router definition. The feature remains experimental.
Model comparisons also become easier to audit. Microsoft says comparisons now remain beneath their originating chat in the sessions list. Hovering over a comparison exposes its repository, isolation mode, base branch, attempt models, effort levels, judge, and synthesizer.
Why it matters
Model routers convert model selection from a per-prompt choice into a repeatable policy. That can reduce manual selection, but the policy lives in a user-authored file and depends on a decision model.
For engineering leaders, routing should be treated as configuration under review. Teams should define which tasks may use which models, validate the resulting selections, and decide whether router files belong in endpoint configuration management.
Comparison metadata improves reproducibility. A result without its model, effort, branch, judge, and isolation settings is difficult to interpret; keeping those values visible makes internal model evaluations more defensible.
This is not automatic model governance. It is user-configurable routing that your organization must validate.
3. Visual Responses and Automations Become Operational Tools
What it is
VS Code 1.147 adds two opt-in visual response formats to the Agents window:
- HTML previews: Enable
chat.htmlPreview.enabledand request the built-in/html-previewskill for reports, diagrams, mockups, tables, inline SVG, and expandable sections. - Native visuals: Enable experimental
chat.nativeComponents.enabledand use/vscode-uifor metrics, badges, tables, bar charts, and expandable details rendered by VS Code.
Microsoft says HTML previews do not support scripts, forms, external links, or external resources. Users can inspect source markup, open an HTML preview at a larger size, or save it as a self-contained HTML file. Native visuals follow the active theme, including high-contrast themes.
The Automations view also adds guided creation. An agent can help define the intended result, schedule, and execution location, while a manual creation option remains available.
Automations can now receive an end date expressed in days, weeks, or months. Selecting Never removes an existing end date. Automation runs are hidden from session lists by default, with a filter available to include them in the Agents window or editor Chat sessions list.
Why it matters
Visual responses can shorten review time for build comparisons, architecture alternatives, and interface mockups. They also create another rendered output path that teams should include in accessibility, export, and content-review testing.
The absence of scripts, forms, external links, and external resources narrows the HTML preview attack surface described by Microsoft. It does not remove the need to inspect generated markup before distributing exported files.
Automation end dates provide a direct control against forgotten recurring work. Your team should require owners, success criteria, execution locations, and end dates for production-facing automations rather than allowing indefinite background tasks by default.
4. Assisted Permissions Gain User-Defined Risk Context
What it is
Microsoft added chat.permissions.assisted.instructions for Copilot harness sessions using assisted permissions. The setting supplies custom guidance to the LLM reviewer that evaluates tool calls.
A team could instruct the reviewer to request approval before package installation or remote pushes while treating tests and read-only Git commands as lower risk. Microsoft says these instructions override the reviewer’s default guidance but cannot override permission rules, organization policy, or the sandbox. They are not sent to the working agent.
The reviewer now also considers who initiated the work. Requests started by automations or other agents are no longer treated as though the user wrote them directly. User answers to agent questions and plan reviews can also influence later permission decisions, including refusals.
Why it matters
The new setting separates risk guidance from the agent’s task instructions. That is useful because the permission reviewer and the working agent have different responsibilities.
For regulated environments, the hierarchy matters: custom reviewer guidance sits below organization policy, permission rules, and sandbox boundaries. Teams should write these instructions as supplemental risk criteria, not as substitutes for enforceable controls.
Changes apply to sessions that are already open. Administrators should therefore test policy updates against active sessions rather than assuming they affect only newly created work.
5. Agent Context and Layout Controls Expand
What it is
The Agents window can now position Files, Changes, and other Details views to the left, right, above, or below the editor through sessions.experimental.detailsPosition. The experimental sessions.experimental.floatingDetails option overlays the Details panel on the editor and allows users to pin, drag, and resize it.
The Customizations view now displays installed extension details, including the version, publisher, and available commands.
Experimental app-window capture is controlled by chat.appWindowScreenshots.enabled. Microsoft says users can select an external application window, attach its screenshot to chat, and mark it up before sending. Markup supports cropping, arrows, shapes, and text and also works with pasted images or images attached from files.
Inline attachment pills are controlled by chat.inlineAttachments.enabled. GitHub issue and pull request pills now expose their title, status, and description on hover and can open the referenced item. Users can drag pills to reposition them within the same prompt, while sent requests show image thumbnails and preserve image and GitHub pills across reloads.
Microsoft says inline attachments are enabled by default in Insiders and disabled by default in Stable.
Why it matters
Screenshot capture gives agents direct visual context from browsers, simulators, and other applications. That reduces manual description, but it can also expose credentials, customer data, notifications, or unrelated windows.
Your organization should decide whether app-window capture is appropriate before enabling chat.appWindowScreenshots.enabled. Image markup helps developers isolate the relevant region, but it does not replace a data-handling policy.
Inline pills improve prompt readability by keeping references adjacent to instructions. For audit-sensitive work, teams should verify that restored pills and image thumbnails preserve enough context to explain what the agent received.
6. Browser Integration Reaches Local and Remote Workflows
What it is
Browser Use is a preview feature controlled by chat.agentHost.browserUse.enabled. Microsoft says it lets a local Copilot harness session interact with the user’s default Google Chrome or Microsoft Edge browser, including existing authenticated sessions.
The prerequisites are explicit:
- Browser component: The GitHub Copilot browser extension.
- Execution location: A Copilot harness session running on the local machine.
- Runtime: Node.js 22 or later on the PATH.
- Activation: Restart VS Code and run
/plugin enable browser-use. - Diagnostics: Run
/browser-use:doctorwhen troubleshooting.
Users approve site access when prompted and complete sign-in or other sensitive steps themselves.
Remote HTML preview is a separate preview controlled by workbench.browser.enableRemoteProxy. It lets the integrated browser open HTML from SSH, WSL, or Dev Container workspaces without copying the files locally. Microsoft says the page retains its file:// address while scripts, styles, images, and fonts load from the remote workspace. Remote files load only from trusted folders.
The integrated browser can now handle non-localhost links through workbench.externalUriOpeners, using workbench.integratedBrowser for configured host patterns. The new workbench.browser.dimOnOverlay setting can prevent pages from dimming when a tooltip or hover overlaps the browser, although notifications still cause dimming.
Why it matters
Browser Use crosses a significant boundary: the agent can operate in a browser that already contains authenticated sessions. Explicit site approval and manual handling of sensitive steps are therefore operational controls, not minor setup details.
Platform teams should test Browser Use in isolated accounts before permitting it in environments with administrative consoles, customer systems, or production access. The capability is preview software with direct interaction privileges.
Remote HTML proxying removes a file-copy step from remote development. Teams should still review workspace trust, remote resource loading, and network boundaries before enabling workbench.browser.enableRemoteProxy.
7. Sandbox, Accessibility, and Modern UI Controls Advance
What it is
When a sandboxed terminal command fails because it cannot access a file or folder, VS Code 1.147 can retry the command inside the sandbox after the user grants permission. Available scopes are the current session, the current workspace when supported, or the user.
Microsoft states that these choices do not grant permission to execute outside the sandbox. This is an in-sandbox retry path.
The Agents window also receives clearer focus and attention indicators. High-contrast themes use dashed outlines for pointer hover and solid outlines for keyboard focus or expanded states. Sessions requiring input receive an amber indicator in light themes or a left-edge marker in high-contrast themes. The color can be customized through agentSessionNeedsInput.foreground under workbench.colorCustomizations.
With Modern UI enabled, workbench.sizeCustomizations accepts whole-pixel corner radii from 0 to 8 and spacing values from 0 to 16. Compact layouts remain gapless. The experimental workbench.experimental.modernUIEditorTabStyle extends connected or pill styling to horizontal panel and side bar tabs.
Why it matters
Retrying inside the sandbox preserves the execution boundary while addressing a missing file permission. That is materially different from approving an unsandboxed retry.
The three permission scopes have different persistence. Your team should prefer the narrowest scope that completes the task and reserve user-level permissions for reviewed, repeatable requirements.
Accessibility changes deserve regression testing in corporate themes. Custom colors, focus outlines, and attention markers must remain distinguishable for keyboard and high-contrast users.
8. Markdown and Pull Request Workflows Change
What it is
The Markdown editor can now keep tables rendered while users edit them one cell at a time. This removes the need to manually preserve pipes and separator rows during routine table edits.
VS Code 1.147 also renders Markdown details and summary sections as collapsible content. Sections carrying the open attribute begin expanded; other sections begin collapsed.
The GitHub Pull Requests extension reaches version 0.180.0 in the release notes. Microsoft highlights an Approve workflows to run button when GitHub Actions requires maintainer approval, a pull request overview that switches to one column below 1000px rather than 768px, and full-width timeline items below the sidebar.
The experimental githubPullRequests.experimental.agentsOverviewTabs setting divides the Agents window pull request overview into Timeline, Checks/Merge, and Metadata tabs. The Timeline tab displays unresolved conversation counts, while Checks/Merge indicates merge readiness.
Why it matters
Cell-level table editing lowers the risk of breaking Markdown structure during documentation updates. Collapsible sections make long operational documents easier to scan without requiring a separate rendering workflow.
The pull request changes bring approval and merge-state information closer to agent-assisted work. Teams should still preserve human review requirements; better visibility is not delegated approval authority.
The move to a 1000px single-column threshold also affects extension layouts in narrower Agents windows. Extension owners should include that width in interface testing.
9. Enterprise Harness and Shell Governance Expands
What it is
Previously, Microsoft says any Copilot enterprise-managed setting disabled Claude and Codex harnesses because those harnesses do not enforce the managed settings. VS Code 1.147 adds vscode.allowedHarnesses so administrators can permit selected harnesses.
The list accepts lowercase claude and codex. An empty list blocks both, unsupported values are ignored, and omitting the setting preserves the previous behavior under managed settings.
The option is available through file-based or MDM-managed settings. Microsoft says server-managed settings do not support it in this release. Other controls, including vscode.temporaryAllowLocalHarness, continue to apply.
Microsoft also warns that permission does not add enforcement. Claude and Codex sessions run without Copilot’s managed permission, sandbox, MCP, and plugin controls.
Managed shell rules now accept JavaScript regular expressions in permissions.deny, permissions.ask, and permissions.allow. The change applies to Shell(...), Bash(...), and PowerShell(...) rules and supplements exact-text and wildcard matching.
Why it matters
vscode.allowedHarnesses resolves a deployment conflict but introduces a clear governance trade-off. An organization can govern Copilot while allowing Claude or Codex, yet those alternative harnesses do not inherit the same controls.
This is not policy parity. It is an exception mechanism.
Security teams should require an explicit risk decision before adding either harness. File-based and MDM delivery also mean endpoint configuration must be verified independently of server-managed policy.
Regular expressions make shell policies more precise, particularly for alternatives and word boundaries. They also make rules easier to misconfigure. Test deny, ask, and allow expressions against representative commands before deploying them across managed endpoints.
10. Platform Deprecations and Removed Settings Require Action
What it is
Microsoft’s VS Code 1.147 release notes deprecate desktop installations on macOS Ventura, macOS Sonoma, and 32-bit ARM Linux. The entries direct readers to the VS Code FAQ for details.
The release notes do not state a shutdown date or continued-support timeline for these platforms.
Microsoft also removed the experimental AI usage statistics Status Bar feature controlled by editor.aiStats.enabled, citing low usage. The setting no longer provides that widget.
Why it matters
Endpoint teams should inventory the named operating systems and architecture now. Because the release notes do not provide a transition date, migration planning should be based on your organization’s validation and procurement lead times rather than an assumed grace period.
The removed editor.aiStats.enabled option should be deleted from shared settings to avoid configuration that suggests a control still exists. Teams relying on the Status Bar widget must revise their usage-observation process.
11. Performance and Reliability Fixes Reduce Background Cost
What it is
Microsoft’s 1.147 contributor ledger includes fixes across agents, terminals, authentication, notebooks, editors, tests, source control, browsers, tunnels, and extension hosting.
Agent and terminal changes include corrected Codex subagent model attribution in PR #340470, streamed BYOK model output in PR #341851, multiple create-terminal actions in PR #346854, package-root detection in PR #347265, task-marker scoping in PR #346994, and pending-line reporting in PR #346122.
Security and accessibility changes include preventing Entra sign-in tokens from being sent to the marketplace in PR #343689, removing redundant screen-reader output from action lists in PR #339952, and correcting authentication-provider cleanup in PR #345127.
State and concurrency fixes cover browser-tab restoration in PR #346996, TypeScript cache invalidation in PR #347028, debugger refresh ordering in PR #347124, forwarded tunnel tracking in PR #347033, notebook output isolation in PR #347117, stale format-on-type cancellation in PR #347115, configuration retries in PR #347074, working-copy backups in PR #347066, virtual document refreshes in PR #346986, and Quick Search buffering in PR #347031.
Microsoft also lists memory-leak fixes for word highlighting in PR #346673, task providers in PR #338231, untitled editor resolution in PR #346674, test result terminals in PR #347183, SCM repository context keys in PR #346683, emitter warning names in PR #347966, the Markers view in PR #348007, and extension-host tree nodes in PR #346727.
Test Explorer receives reduced rescanning in PR #347176 and batched run updates in PR #346656. Microsoft reports idle CPU usage reductions of up to 75% in PR #338767 and a 44 kB memory reduction in PR #342392.
Other listed changes include notebook kernel race handling in PR #341483, PWA server-path correction in PR #330360, webview cache cleanup in PR #346867, UTF-16 column correction in PR #345715, hidden-window extension testing through --hidden-window in PR #344174, empty find-string handling in PR #334803, mobile-emulation element picking in PR #332183, and null review-comment commit handling in extension PR #9168.
Why it matters
The release contains more than visible agent features. The contributor list shows sustained work on memory retention, stale asynchronous operations, authentication cleanup, and background processing.
The vendor-reported 75% idle CPU reduction is the most operationally significant performance claim, particularly for large managed fleets. It should still be validated against your extension mix, repository size, and agent usage before being translated into capacity or battery-life assumptions.
Security teams should prioritize validation of the Entra token-handling change in PR #343689. Developer-platform teams should focus on task-marker isolation, tunnel tracking, backup separation, and extension-host cleanup because those fixes affect long-running workspaces where small leaks and race conditions accumulate.
The Bottom Line
Pilot agent scheduling deliberately. Enable
sessions.sendLater.enabledandchat.automations.enabledonly where local host availability and delayed execution are understood.Treat browser access as privileged automation. Browser Use can operate inside authenticated Chrome or Edge sessions and should begin with isolated accounts and explicit site-access controls.
Review alternative harness exceptions.
vscode.allowedHarnessespermits Claude or Codex alongside managed Copilot settings, but it does not extend Copilot’s managed controls to those harnesses.Test new shell expressions before deployment. JavaScript regular expressions provide more precise managed permission rules but can create broad matches if they are not validated.
Inventory deprecated endpoints. Microsoft deprecated macOS Ventura, macOS Sonoma, and 32-bit ARM Linux without publishing a transition timeline in the release entry.
Agent-first is not a trend, it is the direction. VS Code 1.147 makes that direction operationally concrete through scheduled execution, persistent state, browser interaction, model routing, and enterprise policy boundaries.
Happy Coding!
