The delivery pipeline that closes the loop.
Autoloop grew out of building intothe.zone and mellow.menu, and it is how we build and run our own products. AI agents work through a ranked backlog around the clock, planning, building, testing and reviewing every change, then staging it automatically. A person approves every production release.
How Autoloop works
- 01 plan ranked backlog
- 02 build AI agent
- 03 test sized to risk
- 04 review AI review · CI
- 05 approve human gate
- 06 ship staged · live
Tests come first
Every bug fix starts with a test that fails, so the fix is proven and the bug can’t quietly return.
Testing sized to risk
Each change gets the depth of testing its risk calls for, and the full suite runs every night to catch anything that slipped through.
Escalates, never guesses
Design decisions, payments, sign-in and destructive data changes go to a person. The loop parks the work and moves on.
79 decisions escalated to a person in a typical month
A human at the release gate
The loop does the work. Nothing reaches production until a person approves it.
> autoloop pass
picked · highest-ranked issue · bugs first
branch · isolated workspace · failing test written
tests · sized to risk · passed
review · AI code review + CI checks passed
staging · deployed · smoke tests passed
> release · awaiting human approval
// one_change_start_to_finish
One change, start to finish.
Who does what, in order. The loop runs this for every change, around the clock; a person steps in for open decisions and approves every release.
- 01The agent takes the highest-ranked issue from the backlog, bugs first.
- 02It writes the spec and plan on its own branch, starting with a failing test.
- 03An open design question, or anything touching payments, sign-in or destructive data, is escalated to a person.
- 04The person decides, and the agent carries on.
- 05The change goes up as a pull request.
- 06Tests sized to the risk of the change run, alongside an AI code review and the CI checks.
- 07If anything fails, it goes back to the agent to fix and retry.
- 08When everything is green, the change merges and deploys to staging.
- 09Staging is smoke-tested.
- 10The release waits for a person.
- 11A person approves the release to production.
- 12Production is monitored for errors as they happen. Each one is fed straight back into the backlog for triage, and the loop starts again.
// across_the_lifecycle
Every stage, one loop.
From the first line of a requirement to a change running in production, the same loop covers the whole software lifecycle, with a person at the decisions that matter.
01 · backlog
Backlog and requirements
Work arrives as ranked issues, with bugs ahead of features. Larger features get a product requirements document: user stories, acceptance criteria and a rollout plan behind a feature flag.
02 · spec
Engineering specifications
Non-trivial changes are designed before they’re built, as a written spec and plan. Open design questions go to a person rather than being guessed.
03 · build
Implementation
Each task runs in its own isolated workspace and branch, so several can be in progress without interfering. Work the loop can’t finish is handed off cleanly, never abandoned.
04 · branch
Source control and branching
One branch per issue, with a consistent naming scheme and commits that reference the issue. Main stays releasable, and when it breaks, fixing it comes first.
05 · pr
Pull requests
Every change goes through a pull request. An AI code review and the full set of CI checks must pass before it merges.
06 · test
Testing
Unit, integration and full-browser system tests, plus policy checks that enforce the codebase’s own rules. Depth is chosen from what a change touches; the whole suite runs every night.
07 · quality
Code coverage and quality
Coverage is measured on every nightly run against a minimum. Linting runs on every change; security and dependency scans run nightly, and a regression opens a new issue.
08 · deploy
CI/CD
Merged work deploys to staging automatically and is smoke-tested there. A person approves each production release, and rollback is a deliberate, one-step decision.
09 · observe
Back to the backlog
Production is monitored for errors as they happen; each is fed back into the backlog and triaged in real time, alongside anything the nightly checks catch. Then the loop starts again.