Visual Studio Code 1.144 moves GitHub Copilot agents beyond the lifetime of an editor window. The October 2026 release combines persistent local and cloud sessions, Agent Host Protocol v1.0, Copilot Fusion, model comparisons, image generation, computer use, connectors, session organization, governance controls, terminal changes, and tokenless extension publishing.
For CTOs and engineering leaders, the signal is clear: agent execution is becoming a managed development environment rather than a chat feature. Your platform team now needs policies for where agents run, which tools they can invoke, what context they retain, and how multi-model usage is billed.
1. Agent Sessions Continue Across Local, Remote, and Cloud Environments
What it is
Microsoft’s VS Code 1.144 release notes list Stable downloads for version 1.144.0 and introduce a separate Agent Host process. The host keeps agent sessions running independently of an individual VS Code window, either on the local machine or a remote system.
Microsoft says the Agents window provides one interface for sessions across workspaces and environments. A developer can start a GitHub-hosted cloud task from Chat or the Agents window, answer approval requests, inspect changes, and reopen the same conversation from the session list.
GitHub’s October 28, 2026 cloud-session announcement says those sessions run repositories and tools in a managed GitHub environment. They remain accessible through other Copilot clients, including GitHub.com and GitHub Mobile, before the developer returns to VS Code.
Cloud sessions require a paid Copilot plan and remain subject to organization policies. They also use a separate workspace, so local uncommitted changes are not copied automatically.
The new Remote Environments dialog consolidates GitHub Environments, Dev Tunnels, WSL distributions, and SSH hosts. Users can inspect status, connect or disconnect, and control which remotes appear in environment pickers.
Why it matters
Session continuity changes the operating model for agent-assisted development. Closing a laptop no longer necessarily stops a task, but the execution location now determines which repository state, credentials, tools, and policies apply.
This is not transparent workspace roaming. It is execution rerouting.
Your team should define when work may move to GitHub infrastructure and when it must remain on a managed endpoint or approved remote host. Developers also need a clear handoff rule for local uncommitted changes before starting a cloud session.
For regulated environments, repository access and approval policies should be validated before cloud sessions are enabled broadly. The convenience gain does not remove the need for environment classification.
2. Cloud Automations and Agent Host Protocol Expand the Platform
What it is
Cloud Automations arrive in preview under chat.automations.cloud.enabled, which is enabled by default according to the VS Code release notes. Automations can run on GitHub infrastructure according to a schedule or in response to repository events, even when the user’s computer is offline.
Teams can create shared workflows that respond to issues or pull requests. Filters and allowed-tool restrictions can narrow when an automation executes and what it may do.
Microsoft also published Agent Host Protocol v1.0, the protocol VS Code uses to communicate with agents on local and remote systems. A WebSocket-based host can be added through Agents: Add Remote Agent Host..., and Microsoft says client libraries are available in six languages.
This release also allows extensions to run and be enabled or disabled inside the Agents window when extensions.experimental.agentsWindowManagement is set to true.
Why it matters
Cloud automations turn agent behavior into repository infrastructure. That makes trigger design, tool authorization, ownership, and change review platform-engineering responsibilities.
A safe rollout should begin with no allowed tools, then add only the capabilities required by each workflow. The release includes a contribution that defaults new cloud automations to no allowed tools, reinforcing least-privilege deployment.
Agent Host Protocol v1.0 creates a separate integration decision. Teams can build custom agent interfaces or connect agent platforms to compatible clients, but a remote host also becomes a persistent service that needs authentication, lifecycle management, logging, and network controls.
Agent-first is not a trend. It is the direction of the VS Code architecture.
3. Copilot Fusion Routes Work Across Multiple Models
What it is
GitHub Copilot Fusion is generally available, according to GitHub’s October 28, 2026 announcement and the VS Code release notes. Instead of requiring a developer to select one model, Fusion chooses models, reasoning effort, and an execution pattern for each request.
A simple task might use one model. A harder task can use one model to draft a solution and additional models to review or critique it.
Users can choose Standard, High, or Max effort. Fusion High and Max are offered on Copilot Pro+, Max, Business, and Enterprise plans.
GitHub bills each model used by Fusion at its standard rate. Paid Copilot plans receive a 30% discount through November 30, 2026.
Model comparisons, also described as Agent Bake-Off, run the same prompt across multiple models or harnesses. The setting sessions.chat.compareAgents.enabled controls the feature, which is enabled by default for eligible new sessions unless a user setting or experiment overrides it.
VS Code 1.144 extends comparisons to folderless chats, empty repositories, and local repositories without an initial commit. Model comparisons are not supported in Dev Containers.
Why it matters
Fusion removes a manual model-selection step, but it does not remove the cost of model choice. A single request can produce multiple separately billed model invocations.
Your FinOps controls should therefore measure requests by execution pattern, not only by prompt count. The November 30, 2026 discount deadline should be treated as a budget-planning date rather than a permanent unit-cost assumption.
Model comparisons provide a practical evaluation mechanism for engineering teams. Run representative internal tasks across approved models, then compare output quality, latency, and review burden before setting organization defaults.
Do not assume Dev Container workflows have parity. They do not support model comparisons in this release.
4. Image Generation and Computer Use Expand Agent Reach
What it is
Image generation enters preview with GPT Image 2.5 Flare as the default image model. Agents can create images, modify an image supplied as chat context, or generate variations.
The setting chat.copilot.imageGeneration.model selects the image model. Setting it to off disables image generation, while chat.copilot.imageGeneration.watermark adds a perceptible “Made with AI” watermark.
Microsoft says image generation requires access to the selected model and consumes AI credits in addition to the credits used by the conversation model.
Computer Use is experimental and controlled by chat.agentHost.computerUse.enabled. It allows a Copilot-harness session running locally on macOS or Windows to view and control desktop applications.
Users activate it through /computer on and stop it with /computer off. GitHub’s October 28, 2026 announcement says each application requires approval, and macOS can require operating-system permissions.
Canvases remain in preview after their introduction in VS Code 1.142. Version 1.144 makes them easier to discover through Agent Customizations, Browse, and the Add Context menu.
Why it matters
Image generation introduces a separate credit stream and a content-governance decision. Your organization should decide whether generated assets require visible watermarking, review, or provenance records before teams use them in production work.
Computer Use carries a broader endpoint risk because the agent can interact with applications outside the editor and terminal. Restrict the experiment to managed macOS or Windows devices, approved applications, and workflows with observable results.
This is not another editor tool call. It is desktop automation initiated through an agent.
Canvases provide a richer surface for reviewing artifacts and agent output. Platform teams should still approve the extensions or plugins that contribute those canvases, particularly when marketplace customizations can execute code.
5. Permissions, Working Sets, and Memory Add Control Layers
What it is
Assisted permissions became generally available on October 26, 2026, according to GitHub. An LLM assesses each tool call and automatically approves calls it considers low risk; higher-risk or indeterminate calls still require explicit approval.
In VS Code, assisted permissions require Folder isolation and may be restricted by organization policy. Restricted Assisted and Allow All options now remain visible but disabled, accompanied by an explanation.
Organization-managed session defaults can specify the model, Copilot Fusion tier, and permission level. Those defaults take precedence over remembered user selections, while unavailable permission modes fall back to Manual.
The experimental working set records file edits made through Copilot-harness agent tools. Developers can use inline Keep and Undo controls for individual edits or manage pending changes above the chat input.
The working set is enabled by default in the editor and can be disabled through chat.editor.workingSet.enabled. Changes made through terminal commands do not appear in it.
Copilot Memory is also enabled by default, subject to GitHub settings and organization policies. It stores repository facts shared with people who can access the repository and user preferences that apply across repositories.
Why it matters
Assisted permissions reduce approval fatigue, but they depend on model-based risk classification. For high-impact repositories, Manual approval remains the conservative baseline until your security team has evaluated representative tool calls.
The working set improves reviewability only for edits made through agent tools. Terminal-mediated changes still require conventional source-control review, so teams should not treat the working set as a complete change ledger.
Copilot Memory creates a data-classification requirement. Repository facts may be shared with other authorized repository users, while preferences can cross repository boundaries.
Your rollout checklist should identify which conventions agents may retain, who can delete stored memories, and whether organization policy permits the service. Memory reduces repeated prompting, but it also makes retained context part of the control plane.
6. Connectors and Customizations Become a Managed Marketplace
What it is
Copilot connectors enter preview as GitHub-managed MCP servers for services such as Slack, Notion, and Microsoft Learn. They avoid local MCP-server setup and can be reused across supported Copilot clients after browser authorization.
Connectors are enabled by default on Copilot plans and work in VS Code only with the Copilot harness. Business and Enterprise administrators must enable Editor preview features, while organization MCP restrictions and service-specific approvals still apply.
The new Browse experience collects skills, tools, plugins, workflows, GitHub Feed content, and enterprise-configured marketplaces. GitHub Feed requires a supported harness and is unavailable in Local sessions.
Custom-agent editing now provides completions, hovers, and validation for models, tools, skills, hooks, and agent-scoped MCP servers. Files without an explicit target use the active customization harness for guidance.
Developers can manage customizations from Chat using /plugin, /mcp, and /skills. Deep links can add a marketplace, install a plugin, or open a filtered customization page, with confirmation required before installation.
Why it matters
Connectors can reduce setup effort, but they increase the number of external systems whose content and tools may enter an agent session. Your organization should map each connector to data-access rules, MCP restrictions, and service-owner approval.
Marketplace plugins require stronger controls because the VS Code release notes warn that plugins can run code on the machine. Confirmation is necessary, but it is not a substitute for source review and allowlisting.
Richer validation can catch configuration mistakes before an agent runs. Platform teams should standardize explicit target values where the same custom agent file moves between harnesses, avoiding diagnostics that change with the active environment.
Treat customization sources as software supply-chain inputs. Discovery is easier in 1.144; trust still has to be established.
7. Session Management Becomes a Multi-Workspace Workflow
What it is
The experimental session composer places the environment choice first, offering This Computer, Cloud, eligible remote hosts, and supported Dev Containers. Workspace-less chat can be enabled through sessions.chat.experimental.workspaceOptOut.
The Copilot harness remains the default for new sessions. It is built on the Copilot SDK, which Microsoft describes as a shared foundation across Copilot products.
The preview setting chat.agentHost.moveChats.enabled lets users move nested chats between compatible sessions while retaining conversation history and working-folder context. Moves require the same agent and provider, and the destination must support multiple chats.
Nested sessions can now use separate workspaces. A frontend chat and backend chat, for example, can remain under one parent session while operating in different folders or worktrees.
Quick chats also support nested sessions with separate temporary folders. Session groups gain custom icons and theme-aware colors.
Voice Mode remains experimental under agents.voice.enabled. With agents.voice.useBYOKVoiceModel, users can provide an OpenAI live-model API key, while agents.voice.sessionAwareness shares a bounded briefing of tracked sessions and recent active-chat context.
Voice approvals support 25 languages, with English as a fallback.
Why it matters
Multi-workspace sessions fit cross-repository work, but they also make execution context less obvious. Reviewers need to know which folder, branch, worktree, agent, and host produced each change.
Workspace-less chats should be separated from repository-specific work. Temporary folders reduce setup friction but can complicate artifact retention and audit trails.
Voice session awareness exposes session titles, response summaries, pending prompt details, selected models, and attachment names to the configured voice mode. Microsoft says attachment contents are excluded, but the remaining briefing can still contain operational context.
For enterprise adoption, voice should remain opt-in until API-key handling, data routing, language behavior, and session-context exposure are approved.
8. Editor, Issue Reporting, and Terminal Workflows Improve
What it is
VS Code 1.144 adds an experimental worktree-cleanup suggestion controlled by chat.agentSessions.sessionStorageCleanupSuggestion.enabled. When a local worktree’s disk has less than 5 GB available, the Agents window can suggest cleanup regardless of worktree age or count.
The new issue reporter is enabled through issueReporter.wizard.enabled. It uses a single-page form, supports screenshots, screen recordings, logs, archives, images, and videos, and opens in its own window by default.
Setting issueReporter.wizard.openLocation to editor moves the reporter into an editor tab. Teams can add their own GitHub repositories in owner/repo format through issueReporter.additionalRepositories.
On Windows, terminal.integrated.windowsReparentConsoleWindow makes dialogs opened by local terminal tools appear in front of VS Code rather than behind it. The behavior requires terminal.integrated.windowsUseConptyDll, which is enabled by default.
The release recap also identifies a rendered Markdown editing experience and a refreshed light and dark interface.
Why it matters
Agent worktrees can consume local storage long after a session ends. The 5 GB threshold provides a useful warning, but platform teams should still define retention and cleanup practices for managed developer devices.
The issue reporter can improve reproducibility by capturing evidence without leaving VS Code. Teams using it with internal repositories should verify that recordings, logs, and archives do not disclose secrets or regulated data.
The Windows terminal change removes a failure mode in which authentication dialogs appeared hidden behind VS Code. If your organization relies on terminal-launched sign-in tools, test the new default before disabling it.
9. Pull Requests and Extension Publishing Receive Security Updates
What it is
The GitHub Pull Requests and Issues extension reaches version 0.174.0. Its Update Stack action can use the GitHub API to update a stacked pull request without requiring a pull request from that stack to be checked out.
The pull-request overview can also suggest creating a stack when branch relationships align with other open pull requests that are not already stacked.
For extension publishers, @vscode/vsce 4.0.2 or later supports vsce publish --oidc. GitHub Actions requests a short-lived OpenID Connect token and exchanges it for a temporary Visual Studio Marketplace credential valid for up to 15 minutes.
Publishers must create a Trusted Publishing policy for the repository and workflow and grant the publishing job id-token: write. If the token exchange fails, vsce does not fall back to a personal access token.
Trusted Publishing currently supports GitHub Actions only. Microsoft recommends that Azure Pipelines continue using Microsoft Entra ID service connections and notes that global Azure DevOps personal access tokens are scheduled for retirement on December 1, 2026.
Why it matters
OIDC removes a long-lived Marketplace credential from GitHub Actions secrets. That reduces secret-rotation work and limits the lifetime of a publishing credential to 15 minutes.
Extension teams should migrate eligible GitHub Actions workflows to Trusted Publishing before the December 1, 2026 Azure DevOps PAT retirement affects adjacent release processes. Azure Pipelines teams should validate their Microsoft Entra ID service connections rather than copying the GitHub Actions configuration.
The pull-request extension changes also reduce local checkout requirements for stacked changes. Teams using stacked pull requests should test version 0.174.0 against existing branch and review conventions before standardizing the workflow.
The Bottom Line
VS Code 1.144 is an agent-platform release. Persistent execution, protocol-level interoperability, model orchestration, connectors, memory, desktop control, and cloud automation move Copilot beyond an editor conversation.
What This Means for Enterprise Teams
Classify execution environments. Decide which repositories may use local, remote, and GitHub-hosted agent sessions.
Set conservative permission defaults. Begin with Manual approval or tightly scoped assisted permissions, Folder isolation, and no tools for new cloud automations.
Budget for orchestration. Measure Copilot Fusion by model invocation and reassess projected costs before the 30% discount ends November 30, 2026.
Govern context and customizations. Review Copilot Memory, connectors, MCP restrictions, marketplace sources, plugins, image generation, and voice session awareness.
Modernize extension publishing. Adopt OIDC with
@vscode/vsce4.0.2or later where GitHub Actions is supported, and prepare for the December 1, 2026 Azure DevOps global PAT retirement.
Happy Coding!