Groom features like a human. Ship them like a factory.
Rational Delivery is how we build software: people decide what is worth building, a factory does the rest. Rational Trains is that factory: multi-agent, with customized CI/CD and cloud runs. Agents build, independent reviewers check, standardized metrics decide what merges.
- 01IdeaYouAn issue, written from anywhere. A phone is fine.
- 02GroomingYouEpic, tasks, acceptance criteria.
- 03BuildFactoryAn agent works in its own container.
- 04Cross-reviewFactoryA different model reviews the change.
- 05TestsFactoryBuild and tests pass, or nothing ships.
- 06PreviewFactoryEvery pull request gets its own link.
- 07ReleaseFactoryBehind a feature flag, on for everyone.
Numbers, not adjectives
One week of one factory, straight from its weekly report. The factory measures itself on every train and writes every number to BigQuery.
Read the weekly report- of professional work per $1,000
- 895 h
- At full list prices. About 30× a team without AI, by the most cautious of three estimates.
- pull requests merged
- 486
- In five and a half weeks, one part-time founder with the factory. 2 of 489 closed without merging.
- real defects caught before merge
- 298
- In one week. Every finding is checked against the code before it counts.
- per confirmed defect
- $1.11
- What AI review cost per real defect caught before merge. Human review: $32–64.
- line coverage
- 88.2%
- Up 3.4 points in 12 days. Google calls 90% exemplary.
- on 3 of 4 DORA metrics
- Elite
- Against DORA 2024 thresholds. Time to restore: High.
As of 18 October 2026Based on 195,000 lines of code and more than 30 billion tokens burned.
- Train
- A release. It leaves with its cars and arrives as a new version.
- Cars
- One task each, built in its own container, so work never collides.
- Stations
- The stages every car and every train stops at: build, review, checks, release.
- Schedule
- No ship button to babysit. The train runs its stations and reports to your phone.
No sprawl, no drift
AI coding tools tend to let a codebase sprawl and drift: every change looks fine, and the whole slowly gets worse. Rational Trains is built to hold the line.
How quality holds- Gate
- Standardized metrics decide whether a change merges, not another AI's opinion.
- Ratchet
- Code quality can hold or improve, never slip. A change that makes a metric worse does not merge.
- BigQuery
- Every metric is written to BigQuery: functional progress, code quality and component reuse, over time.
- Learning
- Every train ends with a retrospective, and what it finds goes into how the next train runs.
You
Your time goes to the work only you can do.
- Ideas
- Grooming
- Marketing
- Strategy
The factory
Everything between a groomed task and a release.
- Execution
- Several agents work in parallel, each car in its own container, and notify each other as they go.
- Review
- Independent AI reviewers check every car, and the whole train again before release.
- Checks
- Tests, lint and Prettier run in isolation on every change, locally or in the cloud.
- Delivery
- Customized CI/CD merges the cars, resolves conflicts and releases the train behind a flag.
Why you can trust it
Correctness is checked by runs, not by words.
- Cross-review
- Code written by one agent is reviewed by others before it ships. Different models miss different things.
- Critical bugs
- A critical bug found while a feature is being built is fixed before the train leaves, not filed for later.
- Previews
- You judge the result on a preview link, the way your users will see it, not by reading diffs.
- Flags
- Every change ships behind a feature flag. By default it goes out to everyone and the flag is the emergency brake. In gated mode it reaches only you until you roll it out.
- GitHub
- Backlog in Issues, live progress by train in Projects.
- Google Cloud
- Cloud runs for agents and checks, and BigQuery for every metric.
- ntfy
- Pushes to your phone when a train starts, pauses or is released.
- Your toolchain
- Your own tests, linters and formatters, Prettier included.
Run it where your code lives
Installed as a Claude Code plugin, with a setup wizard for your repo, hosting and analytics.
- Cloud
- Cloud runs on Google Cloud. You set tasks and accept results from your phone.
- Local
- The factory runs on your machine, next to your editor.
- Container
- The factory runs in containers you control.
The waitlist opens soon
We are connecting the form to our analytics first, so no sign-up gets lost. Come back in a little while and you will find it here.