Turn small tracker issues into reviewed pull requests with cloud coding agents
Pick small, well-described issues, write them like a spec with a repro and a test command, and assign them to a cloud agent such as GitHub Copilot's cloud agent, Claude Code in GitHub Actions, Codex cloud, Jules, Cursor or Devin. The agent works on a branch and opens a pull request. You review it like a junior colleague's work and merge it yourself. Teams that report good results put build and test instructions in an AGENTS.md or CLAUDE.md file first and kept tasks small.
Each chosen issue gets a pull request with a passing test run that you review, ask to change, merge or close. You stay in control of every merge.
- Difficulty
- Intermediate
- Time
- 1-2 hours setup, then a review per pull request
- Sources
- 5 · 5 independent
- Updated
- 2026-10-13
- Cost:
- Needs a paid agent plan (Copilot, Claude, ChatGPT with Codex, Jules, Cursor or Devin). GitHub Actions minutes may apply.
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
Paste this into the issue body, or into the optional prompt field when you assign the agent. ## Task [One sentence: what should change and why.] ## Repro [Steps, input, or a failing test name. Example: run `npm test -- src/date.test.ts`, the test "formats ISO dates" fails.] ## Expected behaviour [What correct looks like. Include an example input and output.] ## Scope - Files that are likely involved: [src/date.ts, src/date.test.ts]. - Do not change: [public API signatures, database migrations, .github/workflows/, lockfiles unless the task needs it]. - Keep the change under about 50 lines if you can. If it needs more, stop and explain why in the PR description. ## Definition of done - Add or update a test that fails before your change and passes after it. Do not delete, skip or weaken existing tests. - Run [npm test && npm run lint && npm run typecheck] and paste the result in the PR description. - In the PR description, list what you changed, what you did not do, and anything you are unsure about. - Open the pull request for review. Do not merge it. Do not push to [main].
Human checkpoints: the agent stops and asks you
- Choose which issues go to the agent and rewrite them with a repro, expected behaviour and a test command.
- Read any change to CI or workflow files before you approve the workflow run.
- Review and merge every pull request yourself. Close the ones that are not worth finishing.
The agent pushes code to your repository and can change CI workflow files. Never let it merge. Read any change under .github/workflows before you approve workflow runs.
What you need
- A GitHub repository you can assign agents toaccount
- An AGENTS.md or CLAUDE.md with build, test and lint commandsdata
The .NET team says this file had the biggest effect on their success rate.
- A cloud coding agent planplan
For example Copilot with cloud agent, Claude Code GitHub Actions, Codex cloud, Jules, Cursor or Devin.
- A test suite the agent can run in its cloud environmentdata
Steps
Make the repo buildable by an agent
Write an AGENTS.md or CLAUDE.md with the exact install, build, test and lint commands and the rules the agent must follow. Make sure the agent's cloud environment can download packages. The .NET team reports that early failures came from an agent that could not reach packages or build the repo.
Check: A fresh agent session can install dependencies and run the test suite without your help.
Pick small, well-scoped issues
Start with bug fixes that have a repro, test additions, cleanup and small refactors. Avoid performance work, design decisions and platform code the agent cannot build in its Linux runner.
Check: You have three to five issues that each name the files involved and a test command.
Rewrite each issue as a spec
Use the template in the prompt. Put new information in the issue before you assign it, because some agents do not see later issue comments.
Check: Each issue has a repro, expected behaviour, scope limits and a definition of done.
Assign the issue to the agent
Assign it in the tracker (Copilot takes issues from GitHub and Linear, and Cursor takes them from Linear), mention @claude, or paste it into Codex cloud or Jules. Start with one or two issues, not ten.
Check: The agent acknowledges the task and starts a session or a branch.
Check the pull request before CI runs
Open the diff. If it touches workflow files, lockfiles or anything outside the scope, stop here. Then approve the workflow run if your agent needs that.
Check: The diff stays inside the files you named, and CI runs on the branch.
Review it like a new teammate's work
Check that the new test fails without the change, that no test was deleted or loosened, and that the agent reused existing helpers instead of writing a copy. Ask for changes in a pull request comment.
Check: You can explain why the change is correct, and the test output in the description matches CI.
Merge, fix by hand, or close
Merge good pull requests yourself. If a pull request is close, push the last fix yourself. The .NET team reports a much higher merge rate when humans pushed commits to agent pull requests. Close the rest with a note on what the agent missed.
Check: Every agent pull request is merged or closed within a few days, and none sit open.
Feed lessons back into the instructions file
Add each repeated mistake as a rule in AGENTS.md or CLAUDE.md, such as "reuse toReadableText for labels" or "never edit generated files".
Check: The next pull requests stop repeating the same mistake.
What goes wrong
The agent cannot build or test the repo in its cloud environment, so it guesses.
Fix: Write the setup commands into AGENTS.md or CLAUDE.md and allow the package registries it needs. Do not review a pull request whose description has no test output.
Pull requests arrive faster than anyone can review them.
Fix: Limit how many issues are assigned at once and set a rule to close stale agent pull requests.
Read the lesson →The agent writes a new helper that duplicates an existing one.
Fix: Tell it to search for similar functions first and name the module to reuse. In Birgitta Boeckeler's test, four of six agent runs duplicated logic that already existed.
The agent changes or weakens tests so that they pass.
Fix: Require a new failing test and forbid edits to existing assertions without a reason in the PR description. Check the test diff first.
Read the lesson →The agent keeps working on a task that cannot be done instead of saying it is blocked.
Fix: Give it a time or turn limit and ask it to report blockers. Close the task and rewrite it smaller.
The agent does not see review comments or issue updates you add later.
Fix: Put new information in a pull request comment that mentions the agent, or paste it into the agent's chat.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Blog2026-03-23· GitHub CopilotTen months with Copilot coding agent in dotnet/runtime
Toub reports 878 agent pull requests in dotnet/runtime, with 535 merged. Small cleanup, test and bug fix tasks did best. Performance work and native code did worst. An instructions file and network fixes raised success, and review capacity became the limit.
Stephen Toub · devblogs.microsoft.com - Blog2025-06-04· CodexAutonomous coding agents: A Codex example
Boeckeler gave one small UI task to Codex, Jules and Claude Code in six runs. Codex opened a pull request but could not run the UI tests and broke two existing tests. Only two runs reused the existing helper.
Birgitta Boeckeler · martinfowler.com - Blog2025-01-08· DevinThoughts on a month with Devin
The team gave Devin 20 tasks and report 3 successes, 14 failures and 3 inconclusive results. Small glue tasks worked. It got stuck on impossible tasks and produced code that was hard to read.
Hamel Husain, Isaac Flath and Johno Whitaker · answer.ai - Blog2025-05-22· JulesGoogle Jules first look
The author asked Jules to fix a chat box bug. He hit capacity errors, used up his beta task credits on failed attempts, and reports no successful task.
Robert Matsuoka · hyperdev.matsuoka.com - Forum2025-11-12· CursorIn Linear, Cursor agent doesn't have access to PR comments
The user assigned a Linear issue to Cursor, which opened a pull request. When he asked it to address his review comments, it said it could not see them and asked him to paste them.
laniakealandscape · forum.cursor.com
- Use Copilot cloud agent on GitHubchecked 2026-10-13
- Claude Code GitHub Actionschecked 2026-10-13
- Codex cloudchecked 2026-10-13
- Jules documentationchecked 2026-10-13
- Jules usage limitschecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…