Between 25 September and 14 October, 47 tickets on a B2B wholesaler's Magento platform were built end to end by AI agents. Not suggested, not autocompleted. Picked up from Jira, built on a branch, tested against the real stack, put up as a pull request and moved to review. The client agreed to let me publish the numbers, so here they are, including the ones that do not flatter anybody.
What the pipeline is
Four scheduled jobs run Claude Code against the same Jira board and the same repositories the team uses.
Pre-refinement, twice a day. New tickets get a note with what the code currently does, where the change would go and what is unclear.
Estimation, once a day. Each refined ticket gets an estimate on the team's own scale, with the reasoning.
Code review, every 30 minutes. Every new commit on an open pull request gets reviewed, findings inline.
Development, every hour. The agent takes a ticket that is ready, builds it, runs the checks and opens the pull request.
The development agent runs the same tools a developer does: the local DDEV environment, PHPStan, the project's GrumPHP hooks and a Sonar scan. A ticket is not handed over until those pass. People still do QA, acceptance and the release to production. Nothing reaches customers without a human saying yes.
The numbers
47 tickets built in 19 days: 36 in the PWA, 6 in Magento, 4 in a separate integration app and 1 elsewhere. After the development job moved from office hours to around the clock, it built 9 tickets on a single day, twice.
Where those 47 were on 14 October:
2 closed and 9 ready for production.
5 in acceptance, so 16 of the 47 had passed QA.
29 in test, waiting for QA.
2 still in code review.
Around it: 213 tickets carry a pre-refinement note, 164 an estimate and 129 pull request commits got a review.
What came back
8 tickets went back to the agent after QA or review found something, and it reworked all 8. 8 tickets were handed back to a person at some point, and 7 of those were built in the end. One never was: it needed a decision about the product, not code. One ticket is flagged right now because the agent decided a human should look first. That flag is a feature. The worst thing an agent can do is guess quietly.
The two numbers I cannot give you
The first is cost. The agents run on a flat subscription, so there is no meter. I can tell you it is a fixed monthly amount. I cannot honestly give you a cost per ticket yet.
The second is the comparison. How many of these tickets would the team have finished in the same 19 days without it? Nobody measured that beforehand, and guessing would make this post worse. The honest bottleneck right now is the 29 tickets waiting for QA. Building got faster. Testing did not.
What made it work
The agent works inside the real development environment, not a sandbox, so a passing check means what it means for a developer.
The rules are written down. The repositories carry their own guidelines for the agent, and when a person corrects it, the correction goes into those guidelines rather than into a chat.
Nothing merges itself. The pipeline makes pull requests, and people decide what ships.
The agent is allowed to stop. A ticket it does not understand gets a question, not a guess.
If you are considering this
Start with pre-refinement, not development. A note on every new ticket saying what the code does today is useful on its own, it costs nobody any risk, and it shows you within a week how well an agent understands your codebase. Development comes after that.
If you want to know what this would look like on your own board, get in touch.