Your AI Coding Assistant Just Changed Jobs

For years the deal was simple: you picked an editor, and the AI assistant came with it. Copilot lived in VS Code, Claude Code lived in the terminal, Cursor was the editor that was the assistant. This past week, that deal came apart — in a good way, mostly.

The headline is that the assistant is detaching from the editor. The subscription you already pay for now shows up in places it never did, the tools are learning to work while you’re away, and one editor’s entire model supply nearly vanished overnight. If you write code for a living, or even for fun on a Sunday afternoon, this week changed your toolchain. Here’s the rundown.

Copilot breaks out of VS Code

GitHub made the two biggest moves of the week, back to back. On November 3, Copilot became a selectable agent inside JetBrains Air, JetBrains’ agent environment built on the open Agent Client Protocol — the protocol that lets an IDE treat an AI agent as a swappable component instead of a built-in feature. You pick Copilot from Air’s agent picker, run tasks locally in your JetBrains IDE or in isolated cloud environments through Air Teams, and it all bills against your existing Copilot subscription, not JetBrains AI credits.

A day later, Copilot arrived inside Xcode’s own built-in coding assistant, in public preview starting with Xcode 27.3 beta. You turn it on under Xcode > Settings > Intelligence — no extension required — and it works with every Copilot plan. You’ll need a Mac running macOS Tahoe 26.6 or later. This is separate from the existing GitHub Copilot for Xcode extension, which stays available, so Apple developers now have two paths into Copilot inside the same IDE.

This follows Copilot Fusion going generally available on October 28, GitHub’s model router that picks which models handle each prompt at Standard, High, or Max effort instead of making you choose one yourself. The pattern is unmistakable: GitHub is turning Copilot from a VS Code feature into a subscription that follows you into whatever editor you’re already using.

One honest caveat: GitHub’s Xcode announcement doesn’t say how usage is billed, so don’t assume the preview counts against the same allowance as your editor plugin until you’ve checked your usage page. And it’s a beta-only preview — try it on a side machine, not your shipping toolchain.

Claude Code learns to work while you’re away

Anthropic shipped three Claude Code releases between November 3 and 6, landing on version 2.1.319, and the changes are aimed squarely at unattended runs. When an MCP server runs a tool call as a task, Claude Code now moves it to the background right away and reports the result when it finishes, instead of freezing the whole session on a job that might run twenty minutes. The feature-flag gating is gone too, so Bedrock, Vertex, Foundry, and LLM-gateway setups get it as well.

The same release adds a permissions.denyInsteadOfPrompting setting: a tool call that would normally stop and wait for a permission prompt gets denied immediately and the turn carries on. That sounds minor. It isn’t. It’s the setting that makes a 3 a.m. scheduled agent run actually finish instead of sitting there waiting for a human who is asleep. If you run Claude Code unattended — overnight builds, scheduled refactors, CI jobs — turn it on and pre-approve exactly the tools those runs need.

Cursor’s cutoff became a reprieve

The strangest story of the week is the one that didn’t happen. OpenAI said back in August that it would wind down the contract supplying its models to Cursor after SpaceX acquired Cursor’s parent company Anysphere, with a proposed shutoff date of November 12. The stated reason: OpenAI said it “cannot be confident that SpaceX will use our technology within our terms of service.” Cursor had been frozen on the GPT-5.6 family ever since — no GPT-6, no GPT-6.1 — while OpenAI shipped new models everywhere else.

November 12 arrived. And on that day, OpenAI quietly updated its post: “We’ve temporarily extended Cursor’s access to our models beyond November 12. We will provide updates as available.” No new date, no terms, no explanation. As of this writing, Cursor’s model list still shows GPT-5.6 Sol, Terra, and Luna.

So if you pinned a GPT model as your default in Cursor, it still works — for now. But treat the reprieve as a warning, not a resolution. OpenAI still says it won’t supply future models to Cursor, Cursor CEO Michael Truell has said OpenAI models serve only about 5% of Cursor traffic, and the extension is undated, which means it could end with little notice. The safe move this week: if any of your agents, automations, or team rules pin a GPT model in Cursor, move them to Claude, Gemini, or Grok now, while you can do it calmly. Your own OpenAI API key still works in Cursor for local Chat and Agent, and OpenAI’s own IDE extensions keep working — but those paths don’t cover Auto, Cloud, or Background Agents.

And then there’s the other direction

Not everyone is expanding access. Bloomberg reported on November 2 that Goldman Sachs restricted Claude Code and GitHub Copilot to its roughly 12,000 engineers, pulling access from the bankers and traders who’d been using both tools for months. The bank cited “a variety of factors including productivity and risk management” after what it called a “period of experimentation.” Four months earlier, Goldman’s CIO was publicly praising non-engineer “vibe coding.” The experiment ran, the tools spread past engineering, and the bank pulled them back.

There’s also a security footnote worth your attention: for about 11 hours on October 29, Cursor’s extension marketplace served a backdoored copy of Anthropic’s Claude Code extension — versions 2.1.312 and 2.1.313 on macOS and Windows. Anthropic said a “malicious actor used access to Cursor’s marketplace to publish a tampered copy of the Claude Code extension, which we didn’t build or publish.” If you installed or updated the Claude Code extension from Cursor’s marketplace in that window, check your version and reinstall from Anthropic’s official source.

What to do this week

  • Xcode users on Tahoe: try the Copilot preview in the 27.3 beta on a side project. It’s free to try if you already pay for Copilot — just check your usage page afterward to see how it’s billed.
  • JetBrains users: if your team already pays for Copilot, Air now lets you use it there without a second AI budget. Cloud tasks bill to Copilot, not JetBrains credits.
  • Claude Code users: update to 2.1.319, and if you run anything unattended, set permissions.denyInsteadOfPrompting and pre-approve the tools those runs need.
  • Cursor users: move anything pinned to a GPT model to another model now, while the extension holds. Don’t wait for a final date that may arrive without warning.
  • Everyone: check where your coding assistant actually comes from. A version number in a marketplace listing says nothing about who built the file — install extensions from the vendor, not a mirror.

The week started with your assistant living in one editor and ended with it living in all of them — except in the one place where the model supplier and the owner stopped trusting each other. The tools are getting more portable and more autonomous at the same time. Your job is the same as ever: know which one you’re actually running, and what happens when the company behind it changes its mind.