Has VS Code Lost the AI Era? A 2026 Guide to Cursor, Antigravity, Codex, JetBrains, and Modern Development Environments

Updated October 23, 2026: We checked changes since the August edition against official sources—VS Code's weekly releases and the Copilot harness becoming the default, Cursor's move into The X Subscription, Antigravity's IDE extensions and Gemini 4 Argon, JetBrains Air, Devin Local in Devin Desktop, and model updates in Codex and Claude Code—and added "Changes as of October 2026" to Section 4. The conclusion that VS Code has not lost and is becoming a home base for every agent is unchanged; if anything, other vendors are moving the same way.

1. Conclusion: VS Code Has Not Lost, but It Is Becoming the Home Base for Every Agent Rather Than the Single Best AI Appliance

The first impression varies sharply across AI development environments. Cursor presents codebase indexing, multi-file edits, a dedicated Tab model, and Agent as one integrated experience. Antigravity can move from editing to browser verification in one run, while Codex and Claude Code can take long jobs into separate sessions. Beside products designed around those workflows, VS Code's AI layer can initially look restrained.

That does not mean VS Code is simply inferior. Google Antigravity introduced a Manager Surface separate from its Editor View, where multiple agents could work and return plans, screenshots, and browser recordings as reviewable artifacts. VS Code, by contrast, cannot casually break a huge extension market, enterprise administration, Remote Development, debuggers, and notebooks. A dedicated sports car can turn faster than a public railway station.^[2][5]

By August 2026, however, the conclusion has changed. VS Code is no longer trying to be only the place where Copilot lives. It is moving toward one Agent Sessions layer for Copilot, Claude, Codex, local work, background work, and cloud delegation. Since release 1.109 in January 2026, it has been possible to run Claude and Codex beside Copilot. Current documentation describes parallel sessions, forks, checkpoints, subagents, remote hosts, and the Agent Host Protocol.^[6][7]

October 2026 note: This direction has gone further. Since late August, VS Code has shipped almost weekly (1.135 on August 26, 1.142 on October 14). In the latest release, 1.143, the "Copilot harness," built on the Copilot SDK and running through the Agent Host, became the default for new sessions. The same release adds "Compare models" in preview, which runs one task in parallel across Copilot, Claude, and Codex harnesses and several models, then lets you compare and combine the results. Conversely, when organizational policy requires the Copilot harness, new Claude and Codex sessions cannot be created.^[16] On August 20, Google also released extensions that bring Antigravity's agents into VS Code, Visual Studio, JetBrains, Zed, and Xcode.^[17] The pattern of other vendors' agents boarding VS Code is clearer than it was in August.

The practical answer is therefore:

If you want the fastest path to a satisfying AI-native interaction, a dedicated product such as Cursor or Antigravity can be better. If you need languages, extensions, enterprise controls, remote environments, and agents from several vendors to coexist for years, VS Code has become stronger in 2026.

Usage data shows that VS Code is already the overwhelming market leader. Where it can appear to lose is not overall adoption, but AI-native mindshare and the first ten minutes of delight.

2. VS Code Leads in Usage—A Visible Fork Does Not Mean the Original Has Lost

Stack Overflow's 2025 Developer Survey collected more than 49,000 responses across 177 countries. Among respondents who named development environments, 75.9% selected VS Code. Visual Studio was at 29.0%, IntelliJ IDEA at 27.1%, and Cursor at 17.9%. Stack Overflow noted that Visual Studio and VS Code kept their leading positions for a fourth year and that subscription AI IDEs had not displaced them.^[1]

A pie chart would be misleading. The survey allowed multiple answers. One developer can use IntelliJ at work, VS Code for quick edits, Vim in a terminal, and Cursor for agent tasks. The shares exceed 100% because the environments are not mutually exclusive.

The relationship between VS Code and Cursor may invite a comparison with Chromium and Chrome, but the usage patterns differ. One browser normally leads an individual browsing session, whereas development environments are often divided by role across work, quick edits, terminals, and agent tasks. Cursor also states officially that it is a fork of the open-source VS Code codebase and regularly merges changes from upstream. It competes with VS Code while amplifying VS Code's interaction model, extension assets, and low switching cost.^[2]

VS Code therefore leads as a finished distribution, while its design also lives inside parts of the competition. Its position is closer to the Linux kernel underneath distributions than Chrome on top of Chromium. Winning as a foundation and producing the most exciting finished product are separate contests.

3. Why Does the Same Model Seem Smarter in a Different IDE?

“The same Claude understands me in Cursor but not in IntelliJ.” “The same GPT produces different results in Copilot and Codex.” That is not purely imagined. The model is an engine. The IDE and agent harness provide the map, instruments, chassis, driving rules, and pit crew.

Six harness components commonly change the result:

1. Context collection: Does the agent see only open files, or also dependencies, Git history, diagnostics, instructions, and previous conversations?
2. Search and indexing: Does it use plain text, or combine symbols, types, references, embeddings, and language servers?
3. Tools: Can it only edit files, or also run a terminal, tests, a browser, MCP tools, and cloud environments?
4. The agent loop: Does it answer once, or inspect a failure, revise the work, and try again?
5. Change control: Are there diffs, checkpoints, worktrees, permissions, and exclusions to contain mistakes?
6. Product prompting: System prompts, context compaction, tool definitions, and stopping conditions differ even when the user's words and model name are identical.

Our earlier article, “Why the Same AI Model Yields Different Intelligence,” examines context injection, prompt transformation, and autonomous loops in depth. The central lesson for this article is that IDE selection depends less on the model menu than on what the product shows the model, allows it to do, and uses to detect failure. For the gap between model benchmarks and field performance, see “Why the Benchmark King Breaks Code in the Field.”

A poor IntelliJ-plus-Copilot experience a few months ago does not prove that IntelliJ will always be weak at AI. JetBrains AI Assistant 2026.2 can host Junie, Claude Agent, Codex, and GitHub Copilot, with ACP, MCP, skills, and AGENTS.md support. Conversely, the same product name can feel entirely different depending on plugin version, selected mode, indexing status, permissions, model, and repository structure.^[12]

October 2026 note: On September 22, JetBrains announced JetBrains Air, which brings together "Air in JetBrains IDEs" for directing agents and verifying their work inside the IDE, "Air Teams" for coordinating team workflows, and "Air Governance" (formerly JetBrains Central) for organizational policy, auditing, and cost control. JetBrains states that agentic development will be multi-vendor and that third-party agents from the ACP Registry can run in the same IDEs.^[20] Its own agent, Junie, also kept changing: the free, Mac-only "Junie Local" (August 24), the free "Junie Lite" tier for simpler work (October 2), and stable Windows support for Junie Local (October 16).^[21]

4. Comparing the Main Environments: Tables Work Only When Every Row Uses the Same Ruler

The following table is not a winner-takes-all ranking. It aligns foundation, agent strength, suitable work, and tradeoffs. Desktop readers can scan across columns; on mobile, each product becomes a vertical card. The same information is a table on a wide screen and a list on a narrow one.

It is inaccurate to call every modern editor a VS Code fork. Cursor explicitly is one. Devin Desktop is the renamed Windsurf, and current Devin documentation describes Devin Local and Cascade as its two local agents. Antigravity has a familiar Editor View, but Google's official announcement does not establish a fork. JetBrains, Zed, and Theia use different foundations. A command palette and a left sidebar do not prove shared source code.^[4][14][15]

Changes as of October 2026 (Since the August Edition)

The table above keeps the August framework, with only the Devin Desktop and Antigravity rows revised. Other major changes are summarized below with dates from each vendor's official sources.

Sources:^{[16][22][23][24][17][18][25][26][27][20][21][4][28]}

Seen together, the direction is consistent. VS Code built a place to compare Copilot, Claude, and Codex; JetBrains Air states that multi-vendor is the premise; Devin Desktop accepts third-party agents through ACP; and Google sent Antigravity's agents into other vendors' IDEs. Every product is moving from a dedicated machine for its own agent toward a home base that carries several agents. Cursor is the countertrend: its new plans concentrate included usage on its own side's models such as Grok and bill Claude, GPT, and Gemini at API list price from the start. The more you switch between models, the more you need to estimate costs before choosing a plan.

5. The Right Choice by Language: Look at the Last Mile, Not Just Editor Speed

A table works for comparing products across consistent dimensions. Language recommendations become fragmented in a four-column mobile table, so individual cards are easier to read.

Sakura Editor, Sublime Text, Vim, and Neovim have not disappeared. The Stack Overflow survey reports Vim at 24.3%, Neovim at 14.0%, and Sublime Text at 10.5%. If the agent lives in a terminal or separate desktop workspace, the human editing surface can remain small. A product such as Codex, which can open work in the editor you choose, separates the agent's workplace from the human's manual editing surface. It also helps to distinguish general conversation apps such as ChatGPT Desktop or Claude Desktop from coding agents such as Codex or Claude Code. The former can be an entry point; the latter own repository context, terminal actions, tests, and diffs needed to finish development work.

6. Why Microsoft's Original Looked Behind: Not Because It Is Open Source, but Because It Has More to Protect

The claim that open source allowed competitors to improve VS Code and overtake Microsoft contains some truth. But “open source makes Microsoft weak” reverses the larger effect. Every successful fork further standardizes VS Code interaction patterns and APIs, while Microsoft retains influence over its distribution, Remote Development, WSL, dev containers, notebooks, accessibility, enterprise policy, and extension security.

Three forces made the original look slow:

• Compatibility tax: Thousands of extensions, operating systems, and enterprise policies cannot be broken casually. A dedicated AI editor can change defaults around its AI experience.
• Organizational boundaries: VS Code, GitHub, Copilot, Azure, and Microsoft security requirements cross team boundaries. A focused startup can decide faster.
• Different strategy: Cursor sells one coherent agent experience. Microsoft connects an editor, GitHub, CLI, cloud agents, and third-party agents across a development lifecycle.

VS Code's 2026 Agent Host moves long-running agent work out of the extension host into a separate process, enabling shared clients, remote execution, and work that can continue when no editor is attached. The Agents window spans workspaces and brings Copilot CLI, cloud sessions, and Claude Agent into one view. This is less dramatic than a brilliant one-shot answer, but more important for daily organizational infrastructure.^[8][9]

Weaknesses remain. Discovery is difficult, and a new user must understand Copilot, CLI, Agent Sessions, the Agents window, and extensions. Dedicated products make the initial instruction—“ask here”—much clearer. VS Code's usage display in the status bar is operationally convenient, but a readable limit meter alone does not determine agent quality.

7. Why Mac First? Product Teams Target Early Customers and Smaller Test Matrices, Not Global PC Share

Windows represents much of the general PC market, yet some advanced AI desktop products have launched on Mac first. The sequence makes more sense once the general PC population is separated from the developer population most likely to adopt an early release.

In the 2025 Stack Overflow survey, professional-use OS responses included Windows at 49.5%, macOS at 32.9%, Ubuntu at 27.7%, and WSL at 16.8%. Multiple answers were allowed, but the result shows that macOS is heavily overrepresented among developers compared with the overall computer market. For an early product aimed at advanced developers, Mac can represent roughly one-third of the target group even while it remains a global minority.^[1]

Apple Silicon also reduces the number of common hardware combinations, while Unix-like shells, permissions, and toolchains are relatively consistent. Windows development can involve a native sandbox, PowerShell, WSL1 or WSL2, Windows and Linux file systems, and enterprise security products. The initial test matrix is larger. This is an inference from the published platform and support structure, not a single official statement in which every AI company gives the same reason for launching on Mac.

An experience that Devin Desktop feels comfortable on Mac but heavy on Windows should not become a permanent product verdict, but it should not be dismissed either. Measure memory, indexing time, and concurrent agents on the same repository. Repository placement matters too. For Linux-targeted web work, OpenAI's current WSL guide recommends running Codex inside WSL2, keeping the repository under /home/..., and opening VS Code as a WSL remote window. Keeping the tools and agent on the Linux side removes a boundary that appears when WSL repeatedly scans files on C:.^[11]

October 2026 note: Windows support is also progressing. The Antigravity app added Windows file and network sandboxing in v2.15.1 (September 19), the Codex app shipped a series of WSL terminal and worktree fixes from September into October, and JetBrains declared Junie Local's Windows support stable on October 16.^[19][25][21] Products that started on Mac are best seen as now tightening their Windows and WSL support.

Native Windows remains the correct choice for .NET desktop applications, drivers, and Office add-ins. Xcode remains necessary for the final Swift and iOS build path. An operating system is not a popularity contest; choose it by where the resulting software must ultimately build and run.

8. Final Answer: Separate the Human Cockpit from the Agent Workplace

IDE selection in 2026 is no longer a search for one smartest product. Agents now live across CLIs, desktop workspaces, IDE extensions, clouds, and browsers. The place where a human reads code can differ from the place where an agent performs a long task.

Use this order when deciding:

1. Choose the language's last mile: IntelliJ for Java, Xcode for iOS, or Visual Studio/Rider for .NET desktop—start with the environment that most reliably owns build, debug, and signing.
2. Choose the size of AI work: completion for a few lines, an IDE agent for several files, and Codex, Claude Code, or a background/cloud agent for work measured in tens of minutes.
3. Check recovery: look for diffs, tests, checkpoints, worktrees, browser verification, and permissions.
4. Reduce OS boundaries: align repository, terminal, and agent in WSL2 for Linux work on Windows, but do not force native Windows work into WSL.
5. Compare on the same real task for a week: measure reviewability, regression rate, and recovery—not demo speed.

For an independent web developer, Cursor as the human cockpit and Codex as the long-running worker can be an excellent combination. For enterprise polyglot development, VS Code as the standard home base with Copilot, Claude, or Codex selected per project can be easier to manage. For a large Java system, it is often more rational to keep IntelliJ's semantic tooling and add the agent of your choice than to migrate completely to an AI-native editor.

October 2026 note: For new Cursor customers, the plans introduced on October 12 bill third-party models such as Claude, GPT, and Gemini separately at API list price. Existing Pro and Pro+ subscriptions continue as before.^[24] If Cursor is your cockpit, check first how much your costs depend on which model you mainly use.

VS Code is not always smarter than Cursor. Cursor is not better at everything. And VS Code has not lost. With 75.9% reported usage, the original is no longer merely chasing the best AI sports car. It is building the large station where different vehicles can arrive, be monitored, and exchange passengers under organizational control.

The winning development environment will not be the one that displays the flashiest model name. It will be the one that gives a model the right context and tools, prevents unnecessary changes, catches false success with tests, and returns a diff that a human can confidently adopt.

Research date: Product capabilities and platform support were checked on August 16, 2026. The "October 2026" notes and "Changes as of October 2026" in Section 4 were checked against official sources on October 23, 2026. AI development products change quickly; plans, limits, supported agents, and preview features may change. Recheck official documentation and test the product on your own repository before adoption.

References

1. [1]Stack Overflow, 2025 Developer Survey. The Dev IDE question allowed multiple answers. All-respondent shares include VS Code 75.9%, Visual Studio 29.0%, Notepad++ 27.4%, IntelliJ IDEA 27.1%, Vim 24.3%, and Cursor 17.9%. Professional OS responses include Windows 49.5%, macOS 32.9%, Ubuntu 27.7%, and WSL 16.8%. ↩
2. [2]Cursor, Security. Cursor describes itself as a fork of Microsoft's open-source VS Code codebase and explains its upstream merge cadence and security-default differences. ↩
3. [3]Cursor, iOS & macOS (Swift). The basic workflow uses Cursor for editing and Xcode for building and running. ↩
4. [4]Cognition, “Windsurf is now Devin Desktop” (June 2, 2026), announcing the rename, Agent Command Center, ACP support, and Devin Local as Cascade's successor. Devin Docs, Cascade (updated September 24, 2026) describes Cascade as one of Devin Desktop's two local agents, with new tabs defaulting to Devin Local. ↩
5. [5]Google Developers Blog, “Build with Google Antigravity” (November 20, 2025). ↩
6. [6]VS Code, “Your Home for Multi-Agent Development” (February 5, 2026). ↩
7. [7]VS Code, “Agents”. ↩
8. [8]VS Code, “Agent Host architecture”. ↩
9. [9]VS Code, “Use the Agents window”. ↩
10. [10]OpenAI, Codex IDE extension. ↩
11. [11]OpenAI, Codex on WSL. ↩
12. [12]JetBrains, AI Assistant in JetBrains IDEs, IntelliJ IDEA 2026.2 Help. ↩
13. [13]JetBrains, “The Go Ecosystem in 2025”. ↩
14. [14]Zed, Agents. ↩
15. [15]Eclipse Theia, Theia Coder. ↩
16. [16]VS Code, Release notes 1.143: the Copilot harness as default, Compare models (preview), and a policy requiring the Copilot harness. Release dates confirmed in the release notes for 1.135 (August 26, 2026) through 1.142 (October 14, 2026). ↩
17. [17]Google Antigravity Blog, “Introducing IDE extensions” (August 20, 2026), covering VS Code, Visual Studio, JetBrains (2026.2.1 or later), Zed, and Xcode (27 beta 6 or later). ↩
18. [18]Google Antigravity, Models (checked October 23, 2026) and “Gemini 4 Argon in Google Antigravity” (October 13, 2026). The model table notes that Argon is not available in the Antigravity IDE but is available in the app, CLI, and IDE extensions. ↩
19. [19]Google Antigravity, Changelog, Antigravity app v2.15.1 (September 19, 2026) Windows sandboxing. ↩
20. [20]JetBrains, “JetBrains Air: Building a System of Products for Agentic Software Development” (September 22, 2026). ↩
21. [21]Junie Blog: Junie Local (August 24, 2026), Junie Lite (October 2, 2026), and stable Windows support for Junie Local (October 16, 2026). ↩
22. [22]GitHub Docs, “Indexing repositories for GitHub Copilot”. Copilot Chat and Copilot cloud agent use the semantic code search index. ↩
23. [23]Cursor, Changelog: Projects (September 10, 2026), Remote Control for Local Agents (October 6, 2026), and Cloud Agent Canvases and Environment Improvements (October 16, 2026). ↩
24. [24]Cursor Community Forum, “Introducing The X Subscription” (staff post, October 12, 2026) and Cursor, Models & Pricing (checked October 23, 2026). ↩
25. [25]OpenAI, Changelog: GPT-5.3-Codex-Spark discontinued (September 14, 2026), GPT-6.1 Sol (September 29), GPT-5.5 retirement (October 14), and WSL-related fixes in the Windows app (September 17 and October 7). ↩
26. [26]Anthropic, Claude Code changelog: AGENTS.md support (v2.1.277, September 18, 2026) and Claude Opus 5.5 as the default Opus (v2.1.280, September 22, 2026). ↩
27. [27]Anthropic, Claude Code “What’s new”: /goal (week 20 of 2026) and auto mode as the default for new sessions on Pro, Max, and Team from August 14, 2026 (week 32). ↩
28. [28]Zed, “Introducing Delta” (August 12, 2026) and “Replace PRs with Delta – Now in Public Beta” (September 16, 2026). ↩