# Working with a coding agent > A coding agent is an AI tool, built on a language model, that reads your code, edits files on a branch, and can open a pull request for you to review before you merge. Hand a coding agent one change, and merge it only after the diff and the checks pass. Track: Agents Card: https://buildsparkai.com/coding-agent Plain text: https://buildsparkai.com/coding-agent.md Last checked: 2026-10-20 ## What it is A coding agent is an AI tool, built on a language model, that reads your code, edits files, and can run commands. You hand it one change as an issue, and you stay the one who merges. On GitHub, the changelog of April 1, 2026 calls this product Copilot cloud agent, formerly known as Copilot coding agent. ## Why it matters **Cost.** Hand the agent one change, not a pile of them. A small change takes less time to review. [GitHub's page on Copilot on GitHub.com](https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent) says Copilot can work on one branch at a time and can open one pull request per task, and to break complex tasks into smaller, focused tasks. A second change is a second task. **Quality.** The issue is the prompt, and it says when the change is done. - **Write the issue like a prompt.** [GitHub's best-practices page](https://docs.github.com/en/copilot/tutorials/cloud-agent/get-the-best-results) says it is useful to think of the issue you assign to Copilot as a prompt, and that an ideal task includes complete acceptance criteria plus directions about which files need to be changed. [OpenAI's prompting guide](https://learn.chatgpt.com/docs/prompting) says a useful Codex prompt names the behavior you want, points to the relevant code or reproduction steps, preserves important constraints, and says how to verify the change. - **Ask for a plan first.** OpenAI's prompting guide says to enter /plan so Codex proposes an approach before editing. [GitHub's changelog of April 1, 2026](https://github.blog/changelog/2026-04-01-research-plan-and-code-with-copilot-cloud-agent/) says you can ask Copilot to produce an implementation plan and review the approach before Copilot writes any code. - **Give it the project's rules in AGENTS.md.** [GitHub's changelog of August 28, 2025](https://github.blog/changelog/2025-08-28-copilot-coding-agent-now-supports-agents-md-custom-instructions/) says Copilot coding agent now supports AGENTS.md custom instructions. GitHub's best-practices page lists AGENTS.md for Copilot cloud agent, and [Anthropic's Claude Code overview](https://code.claude.com/docs/en/overview) says Claude Code can read it in place of a CLAUDE.md. - **Other tools do the same handoff.** Claude Code, [Cursor Agent](https://cursor.com/docs/agent/overview), Copilot cloud agent, and Codex are examples. We're not paid by any maker. When we name a tool, it's there as an example, not an endorsement. **Risk.** Passing checks are not a review. GitHub's page on Copilot on GitHub.com says logs do not replace your own review and testing. [GitHub's review page](https://docs.github.com/en/copilot/how-tos/copilot-on-github/use-copilot-agents/review-copilot-output) says that, by default, GitHub Actions workflows will not run automatically when Copilot pushes changes to a pull request, and to be especially alert to proposed changes in the .github/workflows/ directory. Read those files, then allow the workflows to run. [GitHub's changelog of March 13, 2026](https://github.blog/changelog/2026-03-13-optionally-skip-approval-for-copilot-coding-agent-actions-workflows/) says the workflows wait until a person approves them, that this stays the default, and that a repository administrator can skip that approval. If your repository requires approvals, that review page says your approval of a Copilot pull request won't count, and another reviewer must approve it. [GitHub's kick-off page](https://docs.github.com/en/copilot/how-tos/copilot-on-github/use-copilot-agents/kick-off-a-task) says Copilot receives the issue title, description, and existing comments at assignment time. Text you did not write is untrusted. See [Prompt injection](/prompt-injection). GitHub's best-practices page lists production-critical issues and tasks involving security as ones you may choose to work on yourself. OpenAI's prompting guide says that for tasks delegated to the cloud, internet access is off during the agent phase unless you enable it for the environment. ## Try today Pick one small bug in a repository you can change. Write the issue with two lines that say when it is done, and one line that names the files not to touch. Compare it with the issue in the example above. ## Common mistake Merging because the checks pass when the diff also edits a file the issue did not name. GitHub's page on Copilot on GitHub.com says logs do not replace your own review and testing. ## Self-check Which issue is ready to hand to a coding agent? Refunds cite 2025. Done when 10 saved cases cite 2026. ## Practice and deep Example: made up to show the idea. The issue says: Add a ticket-filing tool to the support assistant. Done when the 8 saved eval cases still pass and 3 new cases each file a ticket. Change: ticket.js and ticket.test.js. Do not touch: settings/ or deploy files. The diff has three files: ticket.js (the new tool), ticket.test.js (the new test), and settings/api-key.js (the API key setting). Ask about settings/api-key.js before you merge. It is outside the issue. Now you write the line. The problem is done for you: answers about support hours still say 08:00. Write one line that says when the fix is done, before you open the answer. Show one line that fits Done when 5 saved questions about hours each answer 09:00 to 17:00. It fits because anyone can run it and see pass or fail. "Answers about hours are better" does not: no one can tell when it is done. Self-check A pull request changes 40 files to fix a one-line bug. What do you do? A · Ask for the fix and its test as a small pull request. B · Read through the 40 files line by line, then merge it. C · Accept it, as the description explains each change. ## Sources - [Anthropic, Overview](https://code.claude.com/docs/en/overview) - [Cursor, Cursor Agent](https://cursor.com/docs/agent/overview) - [GitHub Changelog, Copilot coding agent now supports AGENTS.md custom instructions](https://github.blog/changelog/2025-08-28-copilot-coding-agent-now-supports-agents-md-custom-instructions/) - [GitHub Changelog, Optionally skip approval for Copilot coding agent Actions workflows](https://github.blog/changelog/2026-03-13-optionally-skip-approval-for-copilot-coding-agent-actions-workflows/) - [GitHub Changelog, Research, plan, and code with Copilot cloud agent](https://github.blog/changelog/2026-04-01-research-plan-and-code-with-copilot-cloud-agent/) - [GitHub Docs, Best practices for using GitHub Copilot to work on tasks](https://docs.github.com/en/copilot/tutorials/cloud-agent/get-the-best-results) - [GitHub Docs, GitHub Copilot on GitHub.com](https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent) - [GitHub Docs, Kick off a task with Copilot agents on GitHub](https://docs.github.com/en/copilot/how-tos/copilot-on-github/use-copilot-agents/kick-off-a-task) - [GitHub Docs, Review output from Copilot](https://docs.github.com/en/copilot/how-tos/copilot-on-github/use-copilot-agents/review-copilot-output) - [OpenAI, Prompting](https://learn.chatgpt.com/docs/prompting) Opened 2026-10-20. GitHub's changelog of April 1, 2026 names Copilot cloud agent. The AGENTS.md changelog is dated August 28, 2025. The approval-setting changelog is dated March 13, 2026. Anthropic's page is the Claude Code overview. Cursor's page is titled Cursor Agent.