Guide

GitHub Copilot in the cloud: a guide to its coding agent

By Atlas AI Resources · Published October 30, 2026

GitHub Copilot can now research a repository, plan changes, and open pull requests in the background, from github.com, GitHub Mobile, an IDE, or the Copilot CLI. This page starts with our own rollout checklist, then summarizes what GitHub's docs pages GitHub Copilot on GitHub.com and GitHub Mobile and Risks and mitigations for GitHub Copilot cloud agent say.

1. Our advice: a first safe rollout

One repository, one task type. Start with something easy to check, such as adding missing tests, before you hand over features.
Keep the human merge. Leave branch protection and required approvals on. The agent's draft pull requests need a person to merge them; let that be the real gate, not a prompt.
Read the diff, not just the log. Green checks and a tidy session log are not proof the change is right.
Narrow every automation. Give scheduled or event-driven automations only the tools the task needs, and limit write access to the triggering issue or pull request.
Judge by kept work. Count pull requests merged and not reverted, not sessions started.

2. Per the docs: what it does

Ask questions: explain code, suggest fixes, generate tests; chat can create branches, update files, and create or update issues.
Agentic work: edit files and run tests and linters in a cloud environment, then propose changes or open a pull request.
Code review: request a Copilot review on a pull request or configure automatic reviews.
Agent Merge: "Watch and fix" works on failed GitHub Actions workflows, merge conflicts, or both.
Automations: run a task on a schedule or on repository events, managed from the repository's Agents tab (eligible private and internal repositories).

Entry points include github.com/copilot, the Copilot panel in a file, issue, or pull request, your GitHub dashboard, and a repository's Agents tab. Per the docs, starting October 28, 2026 the new Copilot in the cloud experience is available to paid Copilot plans, subject to organization and enterprise policies; Free plans keep the older Copilot Chat experience.

3. Per the docs: where it runs

You choose an environment: Cloud (a cloud sandbox), GitHub Actions, or Remote (a custom environment). Cloud sandbox sessions can be resumed across clients and span repositories in one organization; GitHub Actions sessions stay on the website, work in a single repository, and support self-hosted runners and custom setup. Chat data is retained for 28 days on GitHub Actions and for the life of the account in cloud sandboxes. Check which environment is selected before you start, since defaults vary by workflow.

4. Per the docs: limits

One branch at a time and one pull request per task.
GitHub Actions sessions stop at 59 minutes; the limit can't be extended.
Only repositories hosted on GitHub. Rulesets or branch protection can block it unless an administrator adds a bypass.
Without write access, your session is read-only.
It "can produce incorrect or suboptimal code, including code that contains security vulnerabilities." Review and test before production.

5. Per the docs: built-in guardrails

Only users with write access can trigger it; comments from others are never shown to the agent.
It pushes only to its own copilot/ branch (or the pull request branch it was asked to work on) and can't run git push directly.
It can't mark its draft pull requests ready, approve, or merge them, and the person who asked can't approve that pull request.
Workflows don't run on its pull requests until someone with write access clicks Approve and run workflows (configurable).
It checks its own code with CodeQL, the GitHub Advisory Database for new dependencies, secret scanning, and Copilot code review.
Internet access is restricted by a firewall; hidden text such as HTML comments in issues is filtered out before it reaches the agent.
Commits are signed and co-authored by the person who started the task, with a link to the session log.
Automations ignore events from users without write access by default.

6. How this compares

Cursor's cloud agents follow a similar pattern of isolated environments and pull requests for human review; see our Cursor Cloud Agents guide. The same rules apply to any coding agent: enforce review with repository settings, not instructions.