Source: GitHub Copilot cloud (coding) agent docs: overview and risks

GitHub's docs for its asynchronous Copilot agent: Actions-based ephemeral environments, firewall, copilot/ branches, human-only PR approval and other mitigations.

Contents

Summary

Copilot’s asynchronous agent researches a repository, plans, and implements changes in a cloud environment, then delivers a pull request for human review. The docs now call it the “cloud agent”. Safety comes mostly from GitHub’s existing controls: branches, required reviews, workflow approval and a network firewall.

Key claims

  • Started from the Copilot home page, panels on files/issues/PRs, the Agents tab, security alerts, or automations (scheduled or event-triggered).
  • GitHub Actions sessions are limited to one repository and 59 minutes; cloud sandboxes can span repos in one org and resume across clients, but lack custom compute, Actions-specific setup steps and self-hosted runners. Environment setup can be pre-scripted (.github/workflows/copilot-setup-steps.yml).
  • One branch per session, one PR per task; new work goes to a copilot/ branch. The agent can’t run raw git push, mark a PR ready, approve or merge; the requester can’t approve their own agent PR.
  • Actions workflows on agent PRs don’t run until a user with write access approves them (optional auto-run).
  • Only users with write access can trigger the agent; hidden characters (e.g. HTML comments) are filtered to reduce prompt injection; internet access is firewalled by default.
  • The agent self-checks output with CodeQL, secret scanning, dependency advisories and Copilot code review before finishing.
  • Commits are signed, authored by Copilot with the requester as co-author, and link to a session log.
  • MCP servers are supported; by default the GitHub MCP server only sees the working repository.

Feeds

Coding agents

Suggest changes →