AI software engineer: an agent that takes issues and opens pull requests

What an AI software engineer does

It works like a junior-to-mid engineer who never forgets the checklist: small changes, tests, a clear pull request.

• Takes the next issue that's ready, or the one assigned to it.
• Works on its own branch and checkout, so it never touches your working copy.
• Writes or updates tests and runs the checks before asking for review.
• Opens a pull request that explains the change and how it was tested.
• Answers review comments and fixes what's asked.

Goals to give it

Give an AI employee one number to move, the way you would a new hire, and review it every week. Pick one of these to start.

Its daily routine

• Check its open pull requests for review comments and fix those first.
• Take the next issue that's ready; ask on the issue if it's unclear.
• Branch, change, test and run the checks locally.
• Open the pull request and move the issue to review.
• Repeat until working hours end, then post what's in flight.

How to set up an AI software engineer with Claude Code or Codex

1. Write the role brief

   Save it as CLAUDE.md (Claude Code) or AGENTS.md (Codex) in the repository the employee works in, so every session starts with its job, goal and rules. See the CLAUDE.md guide.

   # Role: Software engineer
   Goal: 10 merged pull requests a week.
   
   ## For every issue
   1. Read the issue; ask on it if anything is unclear.
   2. Work on a new branch in your own checkout.
   3. Write or update tests; run lint, type checks and tests.
   4. Open a PR: what changed, why, how you tested it.
   5. Answer review comments before taking new work.
   
   ## Rules
   - Never push to main or merge your own PR.
   - Ask before database migrations, dependency upgrades or
     deleting code you didn't write.

2. Give it the tools it needs

   The repository and its issue tracker, gh or your Git host's CLI, the project's test and lint commands, and a separate checkout or git worktree per task; see the Claude Code worktree guide.

3. Decide what needs your yes

   Merging, deploying, database migrations, dependency upgrades and anything touching secrets or production data should wait for a person.

4. Put it on a schedule

   A role needs a rhythm, not a single prompt. Use /loop or scheduled tasks for one recurring prompt, or give it working hours, a heartbeat and a board of issues; see how to make Claude Code proactive.

Related guides

For Claude Code vs Codex as the engine, see Claude Code vs Codex. To run several engineers at once, read multiple Claude Code agents and Claude Code background agents.

Hire an AI software engineer instead of scripting one

Crewmark has an AI software engineer ready to hire. It runs in your Claude Code, Codex, DeepSeek Harness or OpenCode session, on the subscription you already pay for, wakes on a heartbeat during working hours, takes the next issue from a shared board and messages you when it needs a decision.

1. Pick the role

   Its skills, goal and schedule come set up. Change the goal to your own number.

2. Connect your project

   Run the setup command in the repository or folder it works in. It gets its own session and checkout.

3. Review, don't babysit

   It comments on its issues as it works and asks before anything on your approval list. Read what an AI employee is.

AI software engineer FAQ