Copilot agent mode vs the coding agent
Agentic AI on GitHub is not one product. The exam expects you to pick the right surface for the job, so learn where each agent runs, who is in the loop and what it produces.
The three surfaces
| Surface | Where it runs | Who drives | Output |
|---|---|---|---|
| Agent mode (IDE) | Your editor (VS Code, Visual Studio, JetBrains) | You, interactively | Edits in your workspace |
| Copilot coding agent | Ephemeral environment powered by GitHub Actions | The agent, asynchronously | Branch and pull request |
| Copilot CLI | Your terminal | You, interactively | Local edits and commands |
Agent mode picks files, edits them, runs terminal commands and tests, reads the errors and tries again. You see each step and approve risky tools.
How the coding agent works
You assign an issue to Copilot, or start a task from the agents panel, Copilot Chat or the CLI. The agent starts a session, creates a branch with the copilot/ prefix, commits as it works and opens a draft pull request. The pull request body shows the progress, and the session log shows its reasoning and tool calls.
When it finishes, it requests your review. You then comment with @copilot to request changes, and it continues on the same pull request. The developer who asked for the task cannot approve the resulting pull request, so a second human review is always required.
Workflows on the agent’s pull request do not run automatically. A person with write access must select Approve and run workflows, which stops the agent from triggering CI unchecked.
Quality of the task
The agent only knows what you give it. A one-line issue such as “Fix the login bug” produces a vague or wrongly scoped pull request. Write the issue like a good ticket: the problem, acceptance criteria, files or areas involved. Repository context comes from custom instructions and from the environment set up for the agent.
What the agent can and cannot do
- It can read the repository, run builds and tests, and use configured MCP tools.
- It works in one repository per task and opens one pull request.
- It has a firewall that limits outbound network access by default.
- It pushes only to its own branches, so branch protection and rulesets still apply.
Exam traps
- Agent mode is not the coding agent. Agent mode is local and interactive; the coding agent is asynchronous and uses Actions.
- The agent does not merge its own work. It opens a draft pull request and a human must review.
- Giving more compute does not fix a vague issue. Fix the issue description first.
- Workflows on agent pull requests need explicit approval to run.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. You want Copilot to work on an issue by itself, in the background, and deliver the result for review. What do you use?
Show answer
Answer: B.The Copilot coding agent is assigned to an issue, runs in a GitHub Actions-powered environment and opens a pull request.
2. Where does the Copilot coding agent push its changes?
Show answer
Answer: B.It works on a copilot/ branch and opens a draft pull request. It never pushes to the default branch.
3. Which statement describes agent mode in the IDE?
Show answer
Answer: B.Agent mode works synchronously in your editor, iterating on files and tools with you in the loop.
4. A coding agent pull request is open and a reviewer wants changes. What is the correct approach?
Show answer
Answer: B.Mentioning @copilot in a pull request comment makes the agent iterate on the same pull request.