What happens when you assign a GitHub issue to Copilot, and how do you change the task once Copilot cloud agent has started?
answer
- an assignee like any teammate
- an Actions-powered ephemeral environment
- a fresh copilot/ branch
- the issue as it stood at assignment
- @copilot on its pull request
basics
~20 sCopilot cloud agent works in an ephemeral GitHub Actions-powered environment on a new copilot/ branch, opens a pull request and requests your review. It reads the issue only at assignment; send later changes as an @copilot comment on its pull request.
solid answer
~50 sAssigning an issue to Copilot (Assignees, then **Copilot**; the docs still label this issue-page path public preview) starts a Copilot cloud agent session. Copilot is sent the issue title, description, the comments that exist at that moment and any optional prompt you add, plus the repository's custom instructions. It works in its own ephemeral development environment powered by GitHub Actions, on a new `copilot/` branch cut from the default branch unless you pick another base, opens a pull request and requests your review when it finishes. Comments added to the issue afterwards are never seen, so new requirements go in a comment mentioning `@copilot` on the pull request, and it pushes the follow-up to that branch. Only people with write access can trigger it, and it is not the IDE's agent mode, which edits your local workspace.
go deeper
Recall the flow: assign the issue to Copilot, it works on a copilot/ branch, opens a pull request and asks you for review. Changes after that go to the pull request with @copilot.
Explain the snapshot: title, description, existing comments, optional prompt and custom instructions are sent once. Know the Actions-powered environment, the single branch and the 59-minute session limit.
Show you would write the issue so it stands alone at assignment, split work to fit one session and one pull request, and route clarifications through @copilot on the pull request rather than the issue.
Weigh which backlog issues suit a snapshot-driven background agent with a 59-minute ceiling, and who in the team owns steering its pull requests so they do not stall waiting for a reviewer.
## What assigning the issue starts **Copilot cloud agent** is the part of GitHub Copilot that works on GitHub.com rather than in your editor. Assigning an issue to Copilot is one of several entry points: you open the issue, click **Assignees**, pick **Copilot**, and a dialog offers an **Optional prompt**, a repository, a base branch, and (where available) a model. The docs mark this issue-assignment path as **public preview**, so its details can still change. Once assigned, Copilot starts a **session**: it researches the repository, makes changes, runs what it can, opens a pull request and finally requests a review from you, which triggers a notification. The same agent can also be started from the agents panel on GitHub.com, from an `@copilot` mention on a pull request, and from integrations such as Slack or Teams. ## What Copilot receives, and what it never sees The issue is a **snapshot**, not a live feed. The docs list exactly what is sent. | Sent at assignment | Not sent | |---|---| | Issue title and description | Comments added to the issue after assignment | | Comments that already exist on the issue | Edits made to the issue later | | Your optional prompt from the dialog | Comments from users without write access, at any time | | The repository's custom instructions | Repositories other than the one it works in | This is the most common surprise: a teammate clarifies the requirement on the issue an hour later, and the pull request ignores it. The fix is to put the clarification where Copilot listens, which is its pull request. ## Where the work happens - **Environment**: an ephemeral development environment **powered by GitHub Actions**, where Copilot can explore the code, build, and run tests and linters. A repository can prepare it with a setup-steps workflow file. - **Branch**: a new branch whose name starts with `copilot/`, based on the default branch unless you choose another base in the dialog. The agent can push only to that one branch. - **Scope**: one repository per task, one branch at a time, and exactly one pull request per assigned task. - **Time**: each session has a hard **59-minute** limit that cannot be extended, so a large issue should be split before it is assigned. - **Identity**: commits are authored by Copilot with the person who assigned the issue as co-author, signed, and linked to the session logs. ## Steering after it starts 1. Open the pull request Copilot raised for the issue. 2. Write a comment that mentions `@copilot` and states the change, for example a new acceptance criterion. 3. Copilot reacts with an eyes emoji, a "Copilot has started work" event appears in the timeline, and a new session pushes commits **directly to that pull request's branch** by default. 4. If you would rather have a separate pull request, say so in the comment; you can also check out the branch and push commits yourself. Copilot remembers context from earlier sessions on the same pull request, so follow-ups are cheaper than starting over. ## Common surprises - **Nothing starts.** The cloud agent must be enabled: on Copilot Business and Copilot Enterprise an administrator enables the policy, and repository owners can opt some or all repositories out. - **The clarification was ignored.** It was written on the issue after assignment; the issue is a snapshot. - **The work landed in another repository.** The dialog's repository dropdown was changed; without a choice Copilot uses the issue's repository. - **The session stopped half-way.** It hit the 59-minute limit; split the issue and assign the parts separately. - **Checks never ran.** Workflows on its pull request wait for a person with write access to approve them. ## Who can start it, and what it is not Only users with **write access** to the repository can trigger the cloud agent, and comments from users without write access are never presented to it. In the assignment dialog you can pick another repository only if you have write access there and the cloud agent is enabled for it. The cloud agent is distinct from **agent mode** in the IDE. Agent mode edits files in your local workspace while you watch; the cloud agent works in the background on GitHub and hands back a pull request for review. Treat its output as a contribution that still needs a human reviewer, because the merge path is deliberately left to people.
- Can an issue assigned to Copilot be worked in another repository or from a non-default branch?Yes. The assignment dialog has dropdowns for the repository and the base branch. It lists repositories you can read, but you can select one only if you have write access and Copilot cloud agent is enabled there. With no choice made, Copilot works in the issue's repository from its default branch, and GitHub warns when the target is in another organization or is public while the issue is private.
- How long can a single Copilot cloud agent session run?Each session has a hard 59-minute limit that cannot be extended or bypassed; a session that exceeds it times out and stops. You can set a shorter limit with `timeout-minutes` in `copilot-setup-steps.yml`, but not a longer one, so large issues should be split into smaller, focused tasks before you assign them.
- A user without write access comments `@copilot` on the agent's pull request. What happens?Nothing. Copilot only responds to comments from people with write access to the repository, and comments from users without write access are never presented to the agent at all. That keeps outside contributors from steering an agent that can push to the repository.
- Copilot keeps reading the issue, so later issue comments update its task.
- The cloud agent edits files in my local editor workspace like agent mode.
- It commits straight to the default branch once its tests pass.
- Anyone who can comment on the issue can start or steer the agent.
- A large issue just runs until it is done; sessions have no time limit.