What Is GitHub Copilot? From Code Completion to a Self-Driving Coding Agent

GitHub Copilot launched in 2021 as a "smart completion that writes the next bit of your code." By 2026 it has become something else. Assign it a single GitHub Issue and walk away, and the AI writes the code in the background, gets the tests passing, opens a pull request (PR), and hands it back — that is "Copilot cloud agent" (called Copilot coding agent until April 2026). It is no longer "a tool that suggests the continuation of what you typed," but closer to "a colleague who takes a task and moves it forward on its own."

Here is the conclusion up front. GitHub Copilot is an AI coding-assistance service from GitHub (owned by Microsoft). As of 2026 there are three ways to use it — (1) code completion (suggests the continuation as you write), (2) chat (ask about code, request fixes), and (3) agent (hand it a task and it implements autonomously). Its defining trait is that it installs as an "extension" into existing editors like VS Code, JetBrains, and Visual Studio, so you can add AI without changing your usual editor. Pricing ranges from a free plan to Pro ($10/mo), Pro+ ($39/mo), and Max ($100/mo), and since June 1, 2026 it has used usage-based billing (AI credits) tied to token consumption.

My stance: for people who "don't want to change the editor they already use," Copilot is one of the easiest ways to start. If your work runs on GitHub Issues and PRs, everything from assignment to review happens inside GitHub. That said, Cursor and Claude Code now also run in the terminal, the IDE and the cloud, so the real differences are less about "where it runs" and more about which models you can use, how usage is counted, and where each one is anchored (I compare all four in Cursor vs Claude Code vs Copilot vs Codex). This article lays out what Copilot can do, the 2026 headliner that is the agent features, pricing, the difference from Cursor / Claude Code, who it fits, and how to start — all with the latest information. For the bigger picture of AI coding, also read what Cursor is and how AI changes the development lifecycle.

1. What Is GitHub Copilot? — From "Completion" to "Self-Driving Agent"

GitHub Copilot is an AI coding-assistance service provided by GitHub (owned by Microsoft). When it launched in 2021, it was an advanced completion tool that "suggested the continuation of the code you were writing, in gray text." Press Tab to accept the suggestion — that experience spread to developers worldwide and became synonymous with AI coding.

But the 2026 Copilot goes far beyond that frame. Beyond completion, you can consult it via chat, and it carries out tasks autonomously as an "agent." In particular, the cloud agent described below: assign it a GitHub Issue, and it reads the repository in the background, writes code, runs tests, prepares a PR, and comes back. From "completion" to "delegation" — that is what Copilot looks like in 2026.

Its defining trait is that it installs as an extension into your existing editor. It can be added to VS Code, JetBrains IDEs (IntelliJ, etc.), Visual Studio, Neovim, and more, so you can add AI while keeping your usual development environment as-is. This is the fundamental difference from Cursor, where you have to switch editors entirely.

2. What It Can Do — Three Modes

In the previous section I wrote "completion, chat, and agent." These three are the pillars of how you use Copilot. Here is what each can concretely do.

Beyond these, there is Copilot code review, which reviews PRs automatically (generally available on paid plans since April 4, 2025). Copilot has also moved beyond the editor: you can use it in the terminal with Copilot CLI (generally available February 25, 2026) and in the desktop GitHub Copilot app (generally available June 17, 2026). Since October 1, 2026, both also offer computer use in public preview, letting Copilot operate desktop apps on screen (macOS and Windows; it asks for approval before controlling an app). From the days when it was "just completion," its reach has extended into every stage of development. For the impact on the whole process, see also how AI changes the SDLC. Sources: GitHub Changelog (April 4, 2025, February 25, 2026, June 17, October 1).

3. The 2026 Headliner: Agent Mode and Cloud Agent

In the previous section I wrote "the 2026 evolution is concentrated in the agent." Understanding this shows what it means for Copilot to have outgrown the "completion tool" label. There are broadly two kinds of agent.

Agent Mode. This agent works inside the editor, in front of your eyes. Ask it to "add this feature," and it edits across multiple files, runs commands when needed, and tries to fix errors itself. In VS Code it began a gradual rollout to all users on April 4, 2025, and on July 16, 2025 it became generally available in JetBrains, Eclipse and Xcode (GitHub Blog, Changelog). You approve and adjust as you watch the results — a structure similar to the "observe → plan → execute → steer" of Claude Cowork.

Copilot cloud agent (called Copilot coding agent until April 2026). This one is distinctive in that it works fully asynchronously in the background. When you assign an Issue to Copilot on GitHub, the agent reads the repository in a temporary environment on GitHub Actions, cuts a branch, writes code, runs tests, and prepares a reviewable PR and notifies you. You just come back to your desk and review the PR. It became generally available on paid plans on September 25, 2025, and since April 1, 2026 you can also have it research, draft an implementation plan, and iterate on a branch before it opens a PR. Since March 17 it automatically uses semantic code search when appropriate, finding related code by meaning rather than exact text matches. Each session runs for at most 59 minutes, and it is available on paid plans only.

Review has changed too. Since March 5, 2026, Copilot code review gathers the related code and directory structure on its own before commenting (an agentic architecture). You can hand its comments to the cloud agent with "Fix with Copilot," and since May 19 you can choose whether the fix goes into the same PR or a new one. Since October 2 you can also request reviews through the REST and GraphQL APIs (Pro and above).

Sources: GitHub Changelog (September 25, 2025, March 5, 2026, March 17, April 1, May 19, October 2) / GitHub Docs: About GitHub Copilot cloud agent

One practical caveat. It is not the case that "throw it an Issue and a perfect PR comes back." The agent is good at quickly drafting, but on vague requirements or thin-context repositories it can produce off-target implementations. Accuracy rises the more you have a good instruction file (such as .github/copilot-instructions.md or AGENTS.md, describing your repo's rules) and a clear Issue. For how to write instruction files, the prompt & instruction-template collection helps.

4. Pricing — Free / Pro $10 / Pro+ $39 / Max $100 and Usage-Based Billing

Pricing as of September 29, 2026 is below (per GitHub's official plans page and docs). Having a free plan is part of Copilot's accessibility.

An important change: on June 1, 2026, billing moved from a "premium requests" model to usage-based billing tied to token consumption (GitHub AI Credits; 1 credit = $0.01). Each plan includes monthly credits: fixed base credits plus a "flex allotment" that GitHub says is designed to adapt as the economics of AI evolve. Credits deplete based on which model you use and how much; when they run out, you can upgrade to a higher plan, set a budget and pay for additional usage, or wait for the monthly reset (the 1st of each month, UTC). Heavy use of expensive models hits the cap faster, so the thinking in token-saving techniques becomes relevant. Verified teachers and maintainers of popular open source may get Pro for free if they meet the eligibility criteria, and verified students can use the free Copilot Student plan (unlimited completions, auto model selection only). Sources: GitHub (plans page, Copilot plans docs, June 1, 2026 changelog).

There is also a change for Business and Enterprise administrators. On September 24, 2026, GitHub added an enterprise- and organization-wide Default policy for new features policy, which takes effect on October 22, 2026. The choices are Enabled, Disabled, Let organizations decide, and from that date any eligible generally available feature an administrator has left unconfigured follows the selected default. Features you have explicitly enabled or disabled keep your choice, and preview features stay opt-in. New models already work this way: they are enabled automatically unless an administrator has turned off the global default or disabled the model (as with GPT-6 Sol and Luna and Claude Sonnet 5.5, added in September 2026). Sources: GitHub Changelog (default enablement, GPT-6 Sol and Luna, Claude Sonnet 5.5).

5. How It Differs From Cursor and Claude Code

"So what's actually different from Cursor or Claude Code?" — a common question. You used to be able to sort them by type — "Copilot is an editor extension, Cursor is an editor, Claude Code is a terminal tool" — but today all three run in nearly every surface: terminal, IDE and cloud. Three differences remain.

Roughly speaking — Copilot "rides the GitHub flow," Cursor means "moving to its editor," and Claude Code is "used through a Claude subscription." Copilot's distinctive color is that picking Copilot as an Issue's assignee starts the work, and review happens on the same screen. On price, Copilot Pro at $10 is half the $20 of Cursor Pro or Claude Pro by monthly fee alone, but it is not the cheapest paid option — ChatGPT Go ($8) also includes OpenAI's coding agent Codex (OpenAI Help). And since each vendor counts usage in different units, the monthly fee can't tell you how much work the same money buys. Conditions and sources are in the four-tool comparison, and Cursor alone is covered in the Cursor deep-dive. Note the three are not mutually exclusive; you can also use them together.

6. Who It Fits — and Who It Doesn't

No tool is universal. Here is where Copilot especially shines, and where another tool fits better.

Personally, if you aren't paying for anything yet, I recommend Copilot as a candidate for "the first AI coding tool to try." It has a free tier, adds straight onto your current editor, and the $10 Pro plan includes the cloud agent. But Pro comes with 1,500 credits a month ($15 worth), so if you hand long tasks to agents every day, keep an eye on usage and consider a higher plan. If you already pay for Claude or ChatGPT, starting with the Claude Code or Codex that comes with it costs nothing extra.

7. How to Get Started

Setup is simple. (1) Enable Copilot with your GitHub account (possible from the free plan; verified students get the free Copilot Student plan, and teachers and maintainers of popular open source may get Pro free if eligible). (2) Sign in from your editor (in VS Code, hover over the Copilot icon in the Status Bar, select "Use AI Features," and sign in with your GitHub account; if you have no plan, you are signed up for Copilot Free — see the official VS Code docs. In JetBrains, Eclipse, Xcode and others, install that editor's Copilot plugin and sign in). (3) Once signed in, completion starts working. Chat and agent are also available from a panel inside the editor.

The trick to getting the most out of it is to put an instruction file in your repository. Copilot reads repository-wide instructions from .github/copilot-instructions.md and path-specific ones from NAME.instructions.md files under .github/instructions/, and its agents also read AGENTS.md (or a CLAUDE.md / GEMINI.md at the repository root) (GitHub Docs). The more rules, commands and conventions they contain, the more accurate it gets. In VS Code, typing /init in chat analyzes your codebase and drafts a starter instruction file. Rather than "dumping a big task on it right away," try small and get a feel for what works — a common starting point for any AI coding tool.

Summary

GitHub Copilot is an AI coding-assistance service from GitHub (owned by Microsoft) that you add to your existing editor as an extension. The three ways to use it are completion, chat, and agent. The 2026 headliner is the agent: "Agent Mode," which autonomously edits multiple files inside the editor, and the "cloud agent," which — given an Issue — auto-generates even the PR in the background. Pricing is Free / Pro $10 / Pro+ $39 / Max $100 (teams $19 / $39 per user), with usage-based billing (AI credits) since June 2026.

Its biggest strength is that "you can add AI without changing your current editor." Compared with Cursor (switch editors entirely) and Claude Code (used through a Claude subscription), it differs in where it is anchored, which models you can use, and how usage is counted. At $10 its monthly entry is on the cheap side, but the cheapest paid option is Codex via ChatGPT Go ($8), and each vendor counts usage in different units. For people not paying for anything yet, it is one candidate to try first, and adding others when it feels limiting is the least-wasteful order.

In the end, what Copilot shows is the very change in development that "AI has moved from 'a tool that completes' to 'a partner you delegate to.'" But a partner you delegate to needs good instructions — set the foundation with the prompt & instruction-template collection before handing work over, and the agent's real power comes out. Also read what Cursor is, how AI changes the SDLC, and token-saving techniques.

FAQ

Q. Can I use GitHub Copilot for free?
A. There is a free plan where you can try completion and chat with a monthly limit. On top of that, verified teachers and maintainers of popular open source may get Pro (worth $10) for free if they meet the criteria, and students get the free Copilot Student plan. GitHub reevaluates eligibility every month (GitHub Docs). Touch it free first, and move to Pro $10 if it's not enough — that's the standard path.

Q. How is Copilot different from ChatGPT?
A. Copilot is an AI "specialized for development and integrated into the editor and GitHub." It reads the context of your code and Issues and does completion, fixes, and even PR creation on the spot. Unlike pasting code into general-purpose ChatGPT to ask, Copilot is built into the flow of development — that's the decisive difference.

Q. What about the copyright and safety of code Copilot writes?
A. Use of generated code is governed by each service's terms. A filter that blocks suggestions matching public code (with code referencing) is available on every plan, including Free. Indemnification against intellectual-property claims (IP indemnity), however, is not included in the individual Free, Pro, Pro+ or Max plans; it comes with Business and Enterprise, and requires the filter to be enabled (GitHub pricing page and FAQ). Still, the principle is not to trust generated code as-is — always review and test it. For handling secrets, see also things to watch when entering AI prompts.

Q. Should I choose Copilot or Cursor?
A. "If you don't want to change your current editor, Copilot; if you're fine moving to an AI-first editor, Cursor." Both let you choose models from multiple vendors and both have free plans; monthly, Copilot Pro is $10 and Cursor Pro is $20. But they count usage differently, so the monthly price alone can't tell you how much work the same money buys. If your work runs on GitHub Issues and PRs, Copilot fits; if repositories on GitLab and elsewhere are in the mix, Cursor is easier. See also the Cursor article.

Q. If I delegate to the agent, do I not need to write code anymore?
A. Not yet to that degree. The cloud agent quickly drafts and handles routine implementation, but requirements definition, design decisions, and final review are human work. It's accurate to see it as a new way of working: "carve out a task and hand it over, then review the PR that comes out." The better your instruction file and the clearer your Issue, the more its power shows.