AgentComparison
GitHub (Microsoft) · product · coding agent on GitHub

GitHub Copilot cloud agent

GitHub's background coding agent that plans, edits a branch and opens pull requests.

Group: Coding agents

GitHub's docs say Copilot cloud agent can research a repository, make an implementation plan and change code on a branch, then let you review the diff, iterate and create a pull request. It works in its own ephemeral environment powered by GitHub Actions, where it can run tests and linters.

Documentation researched, not testedChecked 22 October 2026Vendor documentation, not a test

The decision in plain English

A good fit for bounded tasks in a GitHub repository with clear acceptance tests and a reviewer. Distinguish cloud agent from local IDE agent mode: cloud work runs in an Actions-powered environment and uses both AI credits and Actions minutes. Before enabling it, check repository policy, workflow approval, firewall scope and MCP access. A firewall is not a guarantee that every process is isolated.

Best fit

Teams already on GitHub that can split work into single-repository tasks, review generated changes and run checks before a human merges the pull request.

Poor fit

Cross-repository changes in one task, non-GitHub hosting, or teams expecting the agent to merge changes without a human. Also a poor fit if setup steps or MCP access cannot be reviewed before granting access.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Research, plan and change one GitHub repository in an Actions-powered environment Source · checked 2026-10-22
  • Human PR review and restricted branch writes Source · checked 2026-10-22
  • Default firewall with explicit MCP and setup-step exclusions Source · checked 2026-10-22
  • Default code, dependency and secret checks before completing work Source · checked 2026-10-22

Permissions and human control

  • Only users with repository write access can trigger the agent. It writes to one branch: the existing PR branch for an @copilot PR task, otherwise a new copilot/ branch. Branch protections and required checks still apply.
  • GitHub documents human review and merge: the agent cannot approve or merge a pull request or mark its own draft ready for review. The user who requested the PR cannot supply its required approval.
  • By default, PR Actions workflows wait for a person with write access to click Approve and run workflows. Administrators can allow automatic runs, so confirm the actual repository policy before relying on that gate.
  • The default firewall covers processes the agent starts through its Bash tool inside the Actions appliance. It does not cover MCP servers or configured setup-step processes, and GitHub warns that sophisticated attacks may bypass it. Inspect allowed domains, setup scripts and MCP access separately.
  • Protect Copilot and MCP configuration files with CODEOWNERS and review rules. GitHub distinguishes ordinary Actions secrets from Agents secrets and variables deliberately provided to the cloud agent; grant only the data a task needs.

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Cloud agent requires a paid Copilot plan; the Free plan is not a free cloud-agent tier
  • Uses both GitHub AI credits and GitHub Actions minutes; subscription price alone is not a per-task cost
  • One repository and one branch per task; one pull request; maximum session duration 59 minutes
  • MCP settings can broaden repository context; default firewall does not cover MCP servers or setup steps
  • Branch rules can block the agent; do not remove review requirements merely to make a task run
  • Plans table total credits and Max FAQ base-credit wording differ; variable Flex allotments can change. Verify the actual account allowance.
  • Vendor documentation, not a hands-on test or security audit

Privacy and security

Re-verified 22 October 2026 against GitHub's responsible-use docs for Copilot agents: the cloud agent runs in an ephemeral development environment with a network firewall ENABLED BY DEFAULT that blocks outbound connections to unauthorised hosts against code/data exfiltration; automated security scanning runs on its output; work lands as pull requests for human review and iteration. The doc does not state a prompt/code retention period for the cloud agent. Vendor documentation, not independently audited.

Feature-by-feature: what is documented and what is not

Official sources checked 22 October 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit
RuntimeCloud, not IDE modeAn ephemeral GitHub Actions environment can research, plan and change code. Local IDE agent mode has a different execution environment. Source
Planning entry pointsChannel-dependentResearch/plan/iterate before a PR works on GitHub.com; Teams/Slack support is public preview. Other integrations can create a PR directly. Source
Branch and mergeHuman reviewOnly repository writers can trigger it. It writes one branch and cannot approve, merge or mark its draft ready. The requesting user cannot supply required approval. Source
Workflow approvalDefault can changePR workflows wait for Approve and run workflows by default; administrators can allow automatic runs. Verify the actual policy. Source
Firewall coverageLimited boundaryFirewall covers agent Bash processes in the Actions appliance, not MCP servers or setup-step processes. GitHub warns sophisticated attacks may bypass it. Source
Code checksDefaults, not guaranteesCodeQL, dependency advisory checks and secret scanning are documented defaults that can be disabled. Review output yourself; checks do not prove safe code. Source
Task boundsOne repositoryOne repository, one branch and one PR per task; maximum execution time is 59 minutes. Split larger work without removing review requirements. Source
Usage costsTwo budgetsCloud agent is on paid Copilot plans and consumes AI credits plus Actions minutes. Subscription price is not a per-task total; check both allowances. Source

How to start with GitHub Copilot cloud agent narrowly

  1. Use one repository with a reviewer and clear acceptance tests; check paid-plan and repository policy.
  2. Review setup steps, MCP scope, firewall allowlist and configuration-file ownership.
  3. Set AI-credit and Actions-minute budgets and start one task under the session limit.
  4. Check diff/security results and actual workflow approval policy before a human reviews and merges.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about GitHub Copilot cloud agent

What is GitHub Copilot cloud agent for?

Teams already on GitHub that can split work into single-repository tasks, review generated changes and run checks before a human merges the pull request. The vendor describes it as: GitHub's docs say Copilot cloud agent can research a repository, make an implementation plan and change code on a branch, then let you review the diff, iterate and create a pull request. It works in its own ephemeral environment powered by GitHub Actions, where it can run tests and linters.

Who should look elsewhere?

Cross-repository changes in one task, non-GitHub hosting, or teams expecting the agent to merge changes without a human. Also a poor fit if setup steps or MCP access cannot be reviewed before granting access.

How much does GitHub Copilot cloud agent cost?

Individual subscriptions list Pro $10, Pro+ $39 and Max $100 per user/month. The plans table lists $15 / $70 / $200 total monthly credits, including variable Flex allotments, while its FAQ says Max includes $100/month. Verify actual account allowances rather than treating either as guaranteed. Cloud agent requires a paid plan and also uses Actions minutes. Organization/enterprise additional AI-credit usage is enabled by default unless disabled; this is not an individual-plan assumption.

What can it access and what needs my approval?

Only users with repository write access can trigger the agent. It writes to one branch: the existing PR branch for an @copilot PR task, otherwise a new copilot/ branch. Branch protections and required checks still apply. GitHub documents human review and merge: the agent cannot approve or merge a pull request or mark its own draft ready for review. The user who requested the PR cannot supply its required approval. By default, PR Actions workflows wait for a person with write access to click Approve and run workflows. Administrators can allow automatic runs, so confirm the actual repository policy before relying on that gate. The default firewall covers processes the agent starts through its Bash tool inside the Actions appliance. It does not cover MCP servers or configured setup-step processes, and GitHub warns that sophisticated attacks may bypass it. Inspect allowed domains, setup scripts and MCP access separately. Protect Copilot and MCP configuration files with CODEOWNERS and review rules. GitHub distinguishes ordinary Actions secrets from Agents secrets and variables deliberately provided to the cloud agent; grant only the data a task needs.

What happens to my data?

Re-verified 22 October 2026 against GitHub's responsible-use docs for Copilot agents: the cloud agent runs in an ephemeral development environment with a network firewall ENABLED BY DEFAULT that blocks outbound connections to unauthorised hosts against code/data exfiltration; automated security scanning runs on its output; work lands as pull requests for human review and iteration. The doc does not state a prompt/code retention period for the cloud agent. Vendor documentation, not independently audited.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-10-22. Vendor claims are attributed, not confirmed.

How to evaluate GitHub Copilot cloud agent yourself

  1. Pick one low-risk task that matches its documented capabilities: research, plan and change one github repository in an actions-powered environment.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction