Now that coding agents run terminal commands and call external tools, two settings have outgrown personal preference: which model they use, and which tools switch on by themselves. Both affect a whole team's cost, security, and reproducibility.

GitHub's October 10 update to Copilot for JetBrains addresses exactly this. Enterprises can set the default model through managed settings and disable automatic MCP server startup. There is also a shorter path from a reported diagnostic to a proposed AI fix. Let's take the three apart.

What changed: three new controls

The release condenses to one sentence.

Default model: the org sets it via managed settings (starting model for new chats)
MCP execution: turn off auto-start, switch servers on only when needed (Copilot + Claude sessions)
Diagnostic fixes: a Fix action jumps straight from an error marker to an inline fix request

Enterprise-managed default models: the org picks the starting model

Enterprise administrators can now use managed settings to choose any available Copilot agent model as the default for new conversations. Organizations get more control over the starting model, while a developer's explicit model picker choice is still honored.

Org default (managed-settings.json)
  → new conversations start here
  → a developer switching models in the picker wins for that conversation
  → the next new conversation starts from the org default again

This should feel familiar. A July 1 update first let enterprises pin model to auto so new conversations default to automatic model selection; this release widens the range to any available model. The setting lives at copilot/managed-settings.json in the enterprise's source organization's .github-private repository and applies automatically to users licensed through Copilot Business or Enterprise. You can confirm it is active under Agents and AI controls in enterprise settings.

Why it matters: model choice is now a cost-and-quality policy. A heavyweight reasoning model as the company-wide default quietly inflates token spend; a lightweight-only default drags down agent task success. "The org sets the start, developers pick the exceptions" turns that tradeoff into something a team can manage.

MCP auto-start control: switch tools on only when needed

Second, a new MCP setting lets you disable automatic server startup for Copilot and Claude sessions. Configured tools no longer all activate the moment Copilot starts; you control when they become active.

Auto-start ON (previous default behavior):
  Copilot starts → every configured MCP server boots → tools immediately callable

Auto-start OFF (now possible):
  Copilot starts → servers wait → you start only the ones you need
  → the callable tool surface shrinks

Read alongside the local sandboxing guide, the security angle is clear. If the sandbox governs "how far an executed command may reach," auto-start control governs "which tools end up in the agent's hands in the first place." A tool that starts automatically is a tool the agent can call, and a tool it can call is a tool it will eventually call. This is least privilege applied to the agent inside your IDE.

The direction has been consistent. The September 22 update added a switch for the built-in GitHub MCP server plus persistent per-tool controls, and CLI-side MCP permission and approval policies have kept tightening. "Allow everything" is narrowing to "allow only what you need."

The Fix action for diagnostics: one step from error to fix

Third, and the most tangible change: diagnostic intention menus now include a Fix action. It opens inline chat and asks Copilot to repair the flagged issue, using agent mode when available and ask mode otherwise.

Before: spot error → open chat → explain the error → request a fix
Now:    error marker → Fix → fix proposal right in inline chat

The convenience comes with an obvious caveat. Fix shortens the distance to a proposed fix; it does not judge whether the fix is correct. Agent mode can go as far as editing files, so keep reading the proposed diff and running tests. This connects to the approval flows in the sandbox security guide.

The small print and the deprecation: JetBrains 2025.2+ required

The release also bundles everyday polish: clearer account labels with a Switch Account action, chat messages grouped by turn for faster navigation, copy and edit actions on user messages, plus reliability work across language server startup, model and provider switching, MCP configuration, and worktree flows.

Don't miss the deprecation note:

JetBrains IDE 2025.1 → support ended
Now required: JetBrains IDE 2025.2 or later
(the September 22 release only gave advance upgrade notice for 2025.1)

Check your IDE version first. On 2025.1, none of these features matter because the plugin itself stops working.

CodeBridge Mini Lab: audit your team's defaults

Turn the briefing's action items (check the version and the auto-start setting) into a checklist.

① Environment:
   [ ] JetBrains IDE on 2025.2 or later?
   [ ] Copilot plugin up to date?
② Model default:
   [ ] Is an org-managed default model set? (admins: Agents > AI controls)
   [ ] Can developers still override via the model picker? (verify the escape hatch)
③ MCP surface:
   [ ] With auto-start off, start only the servers you need and work
   [ ] Does the list of running servers match the list you actually use?
④ Fix flow:
   [ ] Hit Fix on one real diagnostic and read the proposed diff
   [ ] If it was an agent-mode edit, confirm tests pass

Fill this in and you get a one-page summary of which model your team's agents start with each day and which tools they hold.

Conclusion: manage agent settings like code

One line to close:

The smarter agents get, the more settings that look like personal taste turn into team policy.

Default models, MCP auto-start control, and the diagnostic Fix action each look small. Together they show the shift: coding-agent competition is moving from "who writes the best code" to "who controls the agent's runtime best." The org sets the starting model for new chats, tools switch on only when needed, and humans read every fix proposal. Those three lines are a decent first draft of your team's AI coding policy for next quarter.

Keep reading

References

Want to learn this with guided, hands-on practice?

If you want to attach Agent and Plan modes to a real project and connect the whole flow through to the runtime environment, applying Copilot to a Java and Spring project follows directly from this article's audit exercise.