The coding agent: a new way to delegate work

GitHub Copilot has moved beyond IDE autocomplete. The new coding agent, built directly into GitHub, operates as an asynchronous teammate: assign it a GitHub Issue, and it will plan the work, open a pull request, write code, run tests, and return the result for your review. You stay in control throughout, deciding what to assign, what to approve, and what to change.

The agent is built on GitHub Actions, giving it a secure and customizable environment to explore the codebase, run tests, and make changes. It can also pull context from related issues and pull request discussions, and it will follow any custom repository instructions your team has configured.

What the agent is good for

The coding agent is ideal for offloading low- to medium-complexity tasks, freeing developers for higher-level problem solving. Common uses include:

  • Adding and extending unit tests, and other repetitive tasks
  • Quick typo fixes, small refactors, and better documentation
  • Bug fixes, UI feature updates, and accessibility improvements

Step one: write and assign the issue

Success starts here. Treat the issue like a briefing for a teammate — the more context, the better the results. Be sure to include:

  • Relevant background: why the task matters and what it touches
  • Expected outcome: what “done” looks like
  • Technical details: file names, functions, or components
  • Formatting and linting rules, especially for custom scripts or auto-generated files

You can assign the issue from github.com, the GitHub Mobile app, or the GitHub CLI. The agent handles well-scoped tasks best but can take on larger ones — it will just take longer. You can also batch-assign multiple issues, which is useful for efforts like broader test coverage.

Some practical tips:

  • Use issue templates with fields for description and acceptance criteria
  • The agent follows repo custom instruction files automatically
  • Images included in the issue are visible to the agent, so you can share visual references for features

Step two: planning begins

Once assigned, the agent reacts with an 👀 emoji and starts a session on GitHub Actions. It opens both a branch and a pull request that will evolve as work proceeds. The agent reads the issue, breaks it down into a checklist of tasks, and posts that checklist to the pull request. Completed items are checked off with each commit, so you can watch the session live, review logs later, or refresh the PR to see how reasoning is progressing.

Step three: coding, testing, committing

After the “Copilot started work” event appears in the PR timeline, the agent will:

  • Modify the codebase based on the issue
  • Run automated tests and linters, updating or generating tests as needed
  • Push commits iteratively as tasks are completed

You can observe in real time and step in if anything looks off, redirecting the agent before it finishes.

Step four: review and merge

When the work is done, the agent tags you for review. You can approve the pull request, leave comments, or request changes. Reviewers are requested automatically based on repo rules, and multiple review cycles are possible until the outcome is what you want.

Once approved, the change follows your normal merge and deploy processes. The agent session ends, but a human can take over from the branch at any time. Note that you cannot assign an issue and then approve the resulting PR yourself — a peer, manager, or designated reviewer must give final approval, ensuring unreviewed code is not merged.

Iterating for better results

The quality of the output is closely tied to the quality of the issue. Clear, scoped, well-documented tasks yield the best outcomes. If the agent gets stuck, tag it in a comment with additional context, and refine the issue requirements as needed. Starting with small tasks — tests, documentation, simple refactors — is a sensible way to get comfortable with the workflow.

“The models we’re using today are the worst ones we’ll ever use — because they’re only getting better,” notes Brittany Ellich, Senior Software Engineer at GitHub. The coding agent is already proving useful in production workflows, and the best way to understand its potential is to try it on your own repositories.