GitHub Copilot Ships a New AI Agent in Visual Studio, Off by Default

GitHub shipped its July 2026 Copilot update for Visual Studio on July 30, and the most telling detail isn't a new feature — it's that Microsoft turned the best part of it off by default. The update adds a new SDK-powered agent to Copilot Chat and a set of built-in "Agent Skills" written by Microsoft's own .NET and Azure teams, but those skills ship disabled, and Microsoft says so explicitly: it's still measuring whether they're worth the tokens they cost.

That's a small, almost bureaucratic-sounding decision. It's also a real signal about where AI coding tools are headed now that usage-based billing has replaced flat "unlimited" plans across the category. Here's what shipped, why it was built this way, and how it stacks up against Cursor and Claude Code.

Quick Take

• What shipped: A new "Agent (Preview)" mode in Copilot Chat built on the GitHub Copilot SDK (the same engine behind the GitHub Copilot CLI), plus built-in Agent Skills for .NET and Azure workflows, inline code review via right-click, org-wide custom instructions, and Git branch context for chat.
• The catch: The built-in skills are off by default. Microsoft says it's evaluating their effectiveness against their token cost before turning them on.
• Why now: GitHub quietly moved Copilot from flat "premium request" limits to per-token AI Credits on June 1, 2026, so every agentic feature now has a visible cost attached.
• Who it affects: Primarily .NET and Azure developers using Visual Studio 2026 (not VS Code); the billing shift behind it affects every Copilot subscriber.

What Actually Shipped

According to GitHub's official changelog for the update, Visual Studio 2026 v18.8 brings a new agent option to the Copilot Chat picker. It's built on the same Copilot SDK that powers the standalone GitHub Copilot CLI, and Microsoft's pitch is narrower and more practical than most agent launches: fewer back-and-forth corrections, shorter responses that are easier to scan, and a more consistent experience as you move between Copilot in Visual Studio, VS Code, and the CLI.

Alongside the new agent, Microsoft added:

• Built-in Agent Skills for ASP.NET Core endpoint creation, .NET performance review, and Azure deployment prep, written by Microsoft's own product teams and surfaced in a new "Built-in" section of the tool picker.
• Inline code review, where selecting code and choosing Copilot Actions → Review Selection returns actionable comments directly in the editor.
• Organization-level custom instructions, so admins can set rules that automatically apply across every repository in the org instead of per-project configuration.
• Git branch context, letting you attach a branch to a chat session for more relevant answers, plus an opt-in fix for C++ projects that pin a toolset version in a separate installation.

None of this requires a new subscription tier — it rolls out through the Visual Studio 2026 Stable Channel to existing Copilot users.

Why the Skills Are Off by Default

This is the part worth paying attention to. In coverage from Visual Studio Magazine, Microsoft explained the decision directly: "Currently, built-in skills are off by default, so you can review and enable only the ones that suit your tasks." The company tied that straight to cost, adding that "as we transition into the new usage-based billing model for Copilot, we want to make sure every token you spend is meaningful."

That's an unusually candid admission from a vendor shipping a new AI feature. Skills that automatically load extra instructions into every agent call also burn extra tokens on every call — and under a metered billing model, an "always-on" default can turn into a surprise line item. Microsoft says it's still gathering evidence on whether the skills measurably improve output before flipping them on for everyone.

The Billing Shift Behind the Decision

None of this happens in a vacuum. GitHub replaced Copilot's old Premium Request Units with per-token AI Credits on June 1, 2026. Under the current structure, Copilot Pro is $10/month with $15 in monthly credits, Pro+ is $39/month with $70 in credits, and a Max tier runs $100/month with $200 in credits. Business seats are $19/user and Enterprise seats are $39/user, and both currently get double promotional credits through August 2026 as GitHub finishes the migration. Ordinary inline autocomplete stays free on every paid plan — only chat, agent mode, code review, and the CLI draw down your credit balance.

Visual Studio had already been feeling the effects of that shift before this release: an earlier July update added a real-time Copilot usage window so developers could see AI Credit consumption before running out mid-month. Shipping new agentic features turned off by default is the same instinct applied one step earlier — don't let people rack up token costs on a feature they haven't chosen to use.

How This Compares to Cursor and Claude Code

GitHub isn't alone in treating agent usage as a cost center to manage carefully. Cursor runs on the same logic: its Pro plan is $20/month, Pro+ is $60, and Ultra is $200, all metered by usage, with a new $120/month "Premium" seat added to Teams plans in July for developers who burn through standard limits. Cursor has also been adding model options — Grok 4.5 landed this month — which, like Copilot's new skills, adds more ways to spend credits faster, not fewer.

Claude Code sits apart from both. It's bundled into a flat Claude Pro or Max subscription rather than metered per agent call, and it's currently the strongest of the three for large, messy codebases — a position reinforced by Claude Opus 5's July 24 launch, which improved coding benchmarks without raising the price. For teams anxious about unpredictable AI bills, that flat-fee model is a real point in Claude Code's favor over token-metered tools like Copilot and Cursor, even if it's less flexible about which model runs under the hood.

If you're still weighing GitHub itself against alternatives for hosting and CI, the platform choice matters too — see our breakdown of GitHub vs. GitLab vs. Bitbucket for that side of the decision.

Who Should Actually Care

If you're not on Visual Studio 2026 — say, you live in VS Code, JetBrains, or a terminal-first workflow — this specific release doesn't touch you directly. But the underlying pattern is worth watching regardless of which tool you use: as coding assistants move from flat subscriptions to metered credits, expect more vendors to start shipping powerful features "off," not "on," until they can justify the token spend. That's a meaningfully different design philosophy than the "everything enabled, figure out the bill later" approach that defined the first wave of AI coding tools. For a broader view of what's worth paying for right now, see our list of developer tools worth using in 2026.

Frequently Asked Questions

Do I need to do anything to get the new Copilot agent in Visual Studio? No immediate action is required for the agent itself — it appears in the Copilot Chat agent picker once you're on Visual Studio 2026 v18.8. The built-in Agent Skills, however, are opt-in: you'll need to enable them individually from the "Built-in" section of the tool picker.

Why did Microsoft ship the new skills turned off? Microsoft says it's still evaluating whether the skills' benefit justifies their token cost under Copilot's new usage-based billing, and wants users to opt in deliberately rather than burn credits on features they didn't choose.

Does this change GitHub Copilot's pricing? No. This is a feature update, not a pricing change. Copilot's move to per-token AI Credits took effect June 1, 2026, and current plans remain Pro ($10/month), Pro+ ($39/month), and Max ($100/month), plus Business ($19/user) and Enterprise ($39/user).

Is this available in VS Code too, or only Visual Studio? This specific release is for Visual Studio (the full .NET/Azure-focused IDE), not VS Code. GitHub ships Copilot updates for VS Code on a separate release cadence.

How does this compare to Cursor or Claude Code? Cursor uses a similar metered-credit model to Copilot's new billing system, while Claude Code is bundled into a flat Claude Pro or Max subscription rather than charged per token — a meaningful difference if predictable costs matter more to you than picking a specific model.

Bottom Line

The new Visual Studio agent itself is a modest, practical upgrade — fewer corrections, shorter answers, tighter integration across Copilot surfaces. The more interesting story is that Microsoft shipped the accompanying skills turned off and said so plainly, tying the decision directly to Copilot's new per-token billing. As AI coding tools across the board move toward metered usage, expect "off by default until proven worth it" to become a more common pattern than the all-features-on defaults of the past few years. For the fuller picture of how Copilot stacks up day to day, read our Cursor vs. GitHub Copilot vs. Claude Code comparison.