From 2 Tickets a Week to 29. Six Weeks After We Installed Nebula Swarm.
Deliveri is a shipping platform for online sellers. We built Nebula Swarm to run its engineering first, and measured what changed from its git, GitHub, Jira and CI history. This is what the numbers show, including the parts that make them look less dramatic.
The First Six Weeks on Nebula Swarm
- 7,673Automated tests added in six weeks. The human team wrote 56 in two years.
- 0Production rollbacks since Nebula Swarm went in.
- 3.3 hMedian time to build a ticket, start to done. The human team took 15.2 hours on the same measure.
- 29.4/wkTickets completed per week, up from 2.0 in the human era.
Source: 37 GitHub repositories, 823 Jira issues and CI logs. "Done" means the code is finished and deployed to dev. How we measured
A Shipping Platform With a Big Roadmap and a Small Team
Deliveri helps online sellers buy carrier labels, connect their Shopify stores and get packages out the door with UPS and USPS. The product is real and complex: carrier integrations, payments, insurance, international duties and a growing internal admin app.
For its first two years a human team built it, with up to six engineers at its peak. By mid-2025 one engineer was making 97% of the commits. The roadmap kept growing. The team didn't.
Three Ways of Building the Same Product
We kept the periods separate on purpose. Plain AI coding tools changed a lot on their own, and we want you to see what Nebula Swarm added on top.
- Apr 2024 – May 2026109 weeks
The human team
Four to six engineers, shrinking to one. Features filed as large stories; almost no automated tests.
- Jun – Aug 202612.6 weeks
Plain Claude Code
No engineers. Two directors driving Claude Code directly, with no harness, no shared process and no reviewer.
- Since Aug 26, 20266.2 weeks
Nebula Swarm
A written plan before any code, Jira tickets in dependency order, agents building each ticket on its own branch, a separate reviewer, and overnight runs.
Same Product, Same Jira, Three Results
| Per week | Human team | Plain Claude Code | Nebula Swarm |
|---|---|---|---|
| Tickets completed | 2.0 | 22.2 | 29.4 |
| Feature and capability tickets | 1.6 | 17.6 | 22.5 |
| Median build time per ticket | 15.2 h | 4.3 h | 3.3 h |
| Automated tests added | 0.5 | ~186 | 1,242 |
Build time is like for like: both eras used the same Jira "In Progress" status (131 human tickets, 134 Nebula Swarm tickets). Test counts are declared test cases on main.
What Nebula Swarm Added on Top of Claude Code
Plain Claude Code got us most of the way past the human team. The harness turned it into a process built for quality, and the clearest gain is in tests. Output and build time improved too.
- 6.7×Automated tests added per week (about 186 → 1,242).
- +32%More tickets completed per week (22.2 → 29.4).
- 23% fasterMedian build time per ticket (4.3 h → 3.3 h).
Plan First. Small Tickets. Nothing Merges Unreviewed.
- Every project starts from a written north star. A PRD and an open-decisions list before any code.
- Work becomes small, ordered tickets. Agents build several in parallel, each on its own branch.
- A separate reviewer checks every change before a human merges it.
- Tests come with every ticket, which is why the test count went from 2,405 to 10,078 on main.
- Work runs overnight. Approved tickets are queued at the end of the day and reviewed in the morning.
- The memory compounds. Conventions and lessons are written down after each task and read before the next one.
182 Tickets, 139 of Them Features
Across the customer app and Genesis, a new internal admin app built almost entirely in this period.
- Genesis admin app: sign-in, orders list with search and filters, order pages, staff actions, USPS pickup scheduling.
- Courier delivery dates: estimated delivery and "delivered on" on every order.
- Security hardening: token-checked APIs, label purchases limited to your own orders, no account enumeration at sign-in.
- Sign-up rework: phone entry with dial codes and a business address step.
- Label fixes customers feel: true 4" × 6" UPS labels, carrier limits checked before payment, protection against buying a second label by accident.
Output in Line With About Ten Engineers
One director spends roughly 35 hours a fortnight directing the crew: writing plans, ruling on decisions, reviewing pull requests. In the last 30 days the agents logged about 699 agent-hours of work.
Measured against how many tickets a competent engineer completes in a week, six weeks of Nebula Swarm output is in line with a team of roughly 10 engineers. That's the low end of our estimate; the method is on the results page.
“Nebula Swarm isn't really about speed. You get much higher-quality output, and that's what lets you build more complex products.”
— Jason Brown, founder of Deliveri [confirm: drafted from Jason's message of 2026-10-28]
Read This Part Too
What we'd tell you before you buy.
- It's six weeks of data. Against two years of the human era, that's a short sample.
- Most of the jump came from AI agents in general. Nebula Swarm added the rest, and the tests. We show both on purpose.
- Nebula Swarm plans work as smaller tickets, especially in Genesis, so ticket counts favor it. Build time per ticket is the fairest single measure.
- Deliveri is our own company, and Nebula Swarm was rolled out by the person who built it. That's why we're taking a small number of founding customers this quarter: to prove the training transfers, measured the same way.
- Not every quality measure improved. The measured part is the tests and the rollbacks. Against plain Claude Code, bug tickets completed a week and reverts plus hotfixes both rose; the results page has the detail.
We'll Record Your Baseline on Day One
On a scoping call we'll look at your team, your roadmap and your current AI tools, and tell you plainly whether Nebula Swarm fits. If it does, we install it in one day on site, measure your numbers on day one, and show you the change at day 15 and day 30.