GitHub Copilot's Coding Agent: Assign an Issue, Come Back to a Pull Request
GitHub Copilot isn't just autocomplete anymore. The coding agent is a fully asynchronous background worker: assign it a GitHub issue, and it independently reads the repository, writes a plan, edits files across your codebase, runs your tests, and opens a pull request for review — no editor session required. This guide covers how it actually works, how it's different from Copilot's in-editor agent mode, and the real limits worth knowing before you hand it a task.
The coding agent works asynchronously in the cloud — you assign an issue and come back to a ready pull request.
Two Different Things Both Called "Copilot Agent"
GitHub Copilot's naming causes real confusion because there are genuinely two different agent experiences bundled under one brand. Agent mode is the synchronous collaborator inside your editor — VS Code, JetBrains, or Visual Studio. You open Copilot Chat, pick "Agent" from the mode dropdown, and it works in real time alongside you: reading files, editing across your codebase, running terminal commands, and fixing test failures while you watch.
The coding agent is different: it's asynchronous and cloud-native. You assign it a GitHub issue — or start a session from any repo URL — and it goes to work independently in the background, on GitHub's own infrastructure. You don't watch it work. You come back later to a pull request with code changes, passing tests, and a self-review already done. This guide focuses on the coding agent specifically, since that's the piece most people mean when they ask "can Copilot just build the thing for me."
From Issue to Pull Request, Step by Step
You assign the task
Assign a GitHub issue directly to Copilot, comment @copilot with instructions, or kick off a session from a repo URL, GitHub.com, an IDE, or the desktop Copilot app. It also integrates with Slack, Teams, Jira, and Linear.
It reads the repo and builds a plan
The agent analyzes the issue description and repository context, then works out which files to create, edit, or delete and what changes to make in each — a plan you can review before it starts, not just a black box.
It branches, codes, and tests — on GitHub Actions runners
Work happens isolated on its own branch. The agent writes the code, runs your test suite, reads any failures, and loops back to fix them without you prompting it again.
It opens a pull request for your review
You come back to a normal PR: a diff, a description of what changed and why, and passing checks. Nothing merges without a human — the agent can't approve or merge its own work.
Each coding agent session runs isolated on its own branch, with the diff surfaced through a normal pull request review.
Where It Still Falls Short
The coding agent can't make cross-repository changes, open several pull requests from one task, sign commits — which matters if your branch protection rules require signed commits — or reach private dependencies that aren't already in the repo. And it can't approve or merge its own pull requests; that restriction is deliberate, not a current limitation waiting to be lifted.
Cost is also a real factor, not just a UX detail. Agent mode and the coding agent both consume premium requests, and the coding agent additionally consumes GitHub Actions minutes since June 2026. A common workflow is to start in ask mode to scope a problem cheaply, then switch to agent mode or hand off to the coding agent only once the task is well defined — because the agent loop's habit of iterating until tests pass means unscoped tasks can burn through a monthly budget fast.
How to Get Started
Copilot Pro ($10/mo)
The cheapest tier that includes coding agent access, alongside a monthly premium-request allowance for agent tasks.
Business and Enterprise
Adds org-level policy controls, MCP server access management, and the ability to assign coding agent tasks at team scale.
Model picker
The Agents panel lets you choose which underlying model handles a given coding agent task, rather than defaulting to one fixed model for everything.
AGENTS.md
Add repo-level instructions the coding agent reads before starting work — coding conventions, test commands, and constraints — to cut down on rewrites.
Bottom Line
The coding agent is the closest GitHub has come to a true background software engineer: assign it a well-scoped issue and it will genuinely produce a reviewable pull request without you sitting there. It's not magic, and it's not a replacement for judgment on ambiguous or cross-cutting work — treat it as a strong option for the narrow, well-defined tickets that pile up in every backlog, and keep agent mode or your own hands on anything that needs real architectural decisions.
More AI Guides, Written Clearly
Browse our full library of AI tool guides, updated regularly as models and features change.
Common Questions
What is GitHub Copilot's coding agent?
It's an asynchronous, cloud-based AI agent that you assign a GitHub issue to. It independently analyzes the issue and repository, writes code, runs tests on GitHub Actions runners, and opens a pull request for human review, without needing an active editor session.
How is the coding agent different from Copilot's agent mode?
Agent mode runs synchronously inside your editor (VS Code, JetBrains, Visual Studio) while you watch and steer it in real time. The coding agent runs asynchronously in the cloud — you assign a task and come back later to a finished pull request.
Can the coding agent merge its own pull requests?
No. This is a deliberate restriction. The coding agent can open a pull request with completed changes, but a human always has to review and merge it — it cannot approve or merge its own work.
How much does GitHub Copilot's coding agent cost?
It requires at least a Copilot Pro subscription, starting around $10 per month, which includes a premium-request allowance. Coding agent tasks also consume GitHub Actions minutes as of June 2026.
More from the creator
Explore two other projects from the same creator.