Key takeaways
- Teams already paying for Copilot often use only the lowest rungs; agent mode and the cloud agent are where the bigger time savings are.
- Repository instructions in .github/copilot-instructions.md, plus path-specific instruction files, give every Copilot surface the same project knowledge.
- The cloud agent is only as good as the issue it receives and the environment it runs in – both can be practised and prepared.
- The cloud agent runs on Linux runners, so iOS builds and signing remain a job for macOS CI and humans.
The Copilot ladder: where most teams stand
GitHub Copilot has grown from an autocomplete into a family of features. It helps to see them as rungs on a ladder, each handing more of the work to the agent:
| Rung | What Copilot does | Who drives |
|---|---|---|
| 1. Completions | Suggests the next lines while you type | Developer, line by line |
| 2. Chat | Answers questions, explains and drafts code on request | Developer, question by question |
| 3. Agent mode in the IDE | Plans a change, edits several files, runs terminal commands and fixes errors | Developer sets the goal and approves steps |
| 4. Cloud agent | Takes an assigned issue, works in its own environment and opens a pull request | Developer writes the issue and reviews the result |
| 5. Code review | Reviews pull requests and suggests fixes | Reviewer decides what to accept |
In our experience most teams sit on rungs one and two. The workshop is built to move them to three and four with guardrails, and to place rung five next to – not instead of – human review. GitHub now calls the agent on rung four the Copilot cloud agent; older documentation and many blog posts still say coding agent.
Half day or full day?
A half day covers rung three: agent mode in the IDE, repository instructions and a team rule for reviewing agent changes. It suits teams that want to try agents without changing their GitHub workflow yet.
A full day adds rungs four and five: writing issues for the cloud agent, preparing its environment, reviewing its pull requests and deciding where automated review helps. This is the better choice when the team already works issue-first on GitHub. For a generic minute plan you can compare against, see the sample agenda.
Talk to an AI Code Engineer about your agentic coding workshop.
In a free call we look at your team, your codebase and your goals and suggest a workshop format that fits – remote or on-site, in English or German.
Block 1: instructions Copilot reads by itself
Copilot picks up repository custom instructions from .github/copilot-instructions.md. More specific instruction files in .github/instructions/ can target parts of the codebase by path pattern, and reusable prompt files hold tasks the team runs often. The cloud agent also reads AGENTS.md files, which is handy if other agents work in the same repository.
The lab: the group writes the repository instructions together, starting with build and test commands and the three conventions reviewers most often have to correct. Then one person per area – backend, web, mobile – writes a path-specific file. We test each one by giving agent mode a task in that area and checking whether it follows the instructions. GitHub's Copilot documentation lists the exact file formats.
Block 2: agent mode on real tickets
Pairs pick a ticket from your backlog and solve it in agent mode, in VS Code or the IDE your team uses. The exercises focus on the moments where people usually go wrong:
- Approving terminal commands. Which commands are routine, which need a second look, which should never run from an agent session.
- Steering mid-task. Stopping the agent when it heads in the wrong direction instead of letting it finish and cleaning up afterwards.
- Tools and MCP servers. Agent mode can use MCP servers; we add one your team needs, such as the issue tracker, and discuss who approves new ones. The MCP workshop goes much further.
- Keeping changes reviewable. Splitting a large change into separate commits or pull requests before it reaches a reviewer.
Block 3: issues the cloud agent can finish
When you assign an issue to Copilot, the cloud agent works in an ephemeral environment powered by GitHub Actions, pushes to a branch and opens a pull request. You steer it by commenting on that pull request. Whether the result is useful depends mostly on the issue. We practise rewriting real issues until they look like this:
Title: Show last sync time on the settings screen Context The app syncs data in the background (see sync/SyncService.kt). Users cannot tell when the last successful sync happened. Expected behaviour - Settings shows "Last synced: [relative time]" under the account section. - If no sync has happened yet, show "Not synced yet". - The value updates when the screen is opened. Done when - A unit test covers the formatting of the relative time. - Existing tests pass with ./gradlew testDebugUnitTest. - No new dependencies.
Issues that only say "add sync info to settings" come back with guesses. Issues with context, behaviour and a definition of done come back as pull requests worth reviewing. Teams that want to formalise this go on to spec-driven development.
Block 4: preparing the cloud agent's environment
The cloud agent needs your dependencies and tools before it can run tests. You prepare them in a workflow file at .github/workflows/copilot-setup-steps.yml with a single job named copilot-setup-steps, which must be on the default branch to take effect. Variables and secrets for the agent live in the repository's copilot environment.
In the lab we write that file for your repository and prove it works with a trivial issue. Two limits come up every time with app teams:
- The agent runs on Ubuntu Linux runners. It can edit Swift code but cannot build or test an iOS app; that stays in a separate workflow on macOS runners – see GitHub Actions for mobile CI with agents.
- Android builds work if the setup steps install the SDK and warm the Gradle cache, but emulator tests are usually better left to your regular CI.
Block 5: review rules and governance
On Copilot Business and Enterprise plans, an administrator has to enable the cloud agent before anyone can use it. That is the moment to set rules, not afterwards:
- Agent pull requests go through the same branch protection, required checks and human approval as any other.
- Copilot code review is an extra reviewer; a person still approves and merges.
- Repositories with production secrets or regulated data get the cloud agent only after a security review.
- Someone owns the instruction files and updates them when reviewers keep correcting the same thing.
The AI code review workshop and the secure agentic coding workshop cover both topics in depth.
For app teams – and running it with us
If your repository holds a mobile app, the release path still needs signing, store listings and review. Copilot app to the App Store and Copilot app to Google Play explain what Copilot can prepare and what stays human. Still deciding between agents? Read Codex vs. GitHub Copilot agent.
Our AI Code Engineers run this workshop remote or on-site, in English or German, inside your GitHub organisation where your policies allow it.
Frequently asked questions
Is the Copilot coding agent the same as the cloud agent?
Yes. GitHub renamed the Copilot coding agent to Copilot cloud agent. Both names describe the agent that works on assigned tasks in a GitHub Actions-powered environment and opens pull requests.
Which Copilot plan do we need?
Agent mode and the cloud agent are available on paid Copilot plans, with usage limits that differ by plan. On Business and Enterprise plans an administrator must enable the cloud agent. Check the current plan details before the workshop.
Can the cloud agent build our iOS app?
No. It runs on Linux runners, so it can change code but cannot run Xcode. Build, test and sign iOS apps in a separate workflow on macOS runners or on a Mac.
Do we need to change how we write issues?
For issues you want to assign to Copilot, yes: context, expected behaviour and a clear definition of done. Many teams find that these issues are also easier for human colleagues.
Free, 15 minutes, no obligation
Plan your agentic coding workshop with us.
Our AI Code Engineers run hands-on workshops on Claude Code, OpenAI Codex, Cursor and AI app builders – and ship apps to the App Store and Google Play every week. Tell us about your team.
AI tools, store policies and requirements change quickly – always check the current documentation of the tools and stores you use. This article is independent advice and not affiliated with or endorsed by Apple, Google, Anthropic, OpenAI or any other company or product mentioned; all trademarks belong to their owners.