The feature pipeline: from user signal to shipped code, behind 1,400 tests
How our ventures ship features with three agents and a test gate: a dreamer that proposes, a groomer that ranks, and an implementer that must pass the same suite a human would.
Feature backlogs rot. Ideas arrive faster than anyone grooms them, tickets duplicate, and the loudest request wins. On our ventures, three agents keep the backlog honest — and a test gate keeps them honest.
The dreamer
Once a day, outside business hours, an agent reads the accumulated signal — GitHub issues, user feedback, operational logs — and drafts candidate feature specs. It archives the signal it consumed, so the same complaint never spawns two candidates. It also refuses to run when the ready queue is already full: eight groomed specs waiting is a signal to build, not to ideate.
The groomer
A second agent dedupes candidates against everything planned, in progress, and shipped; splits oversized ideas into buildable slices; attaches build metadata; and ranks the queue. Grooming is the unglamorous job humans skip when busy — which is exactly why it's the most valuable one to automate. It runs twice a day and queues patiently behind the dreamer instead of racing it.
The implementer
The third agent takes the top ready spec and builds it: code, tests, docs. Then it faces the gate — the full suite, currently 1,400+ checks, plus type checking, linting, and a production build. The gate does not know or care that an agent wrote the code. Fail, and the change does not land. Every iteration appends a receipt to an append-only log: what was attempted, what passed, what shipped.
What the humans do
We read receipts, adjust priorities, and occasionally kill a spec the pipeline loved but the business didn't need. The judgment stayed; the ceremony left.
Honest limitations
The pipeline is only as good as the test suite under it — an untested corner of the codebase is a corner we don't let it touch. Big architectural changes still get planned by people. And roughly one implementation in a handful needs human rework before it lands. We consider that ratio excellent for backlog-class work, and we would not use this machinery for security-critical changes without a human design pass first.
If your product has a backlog nobody grooms, this architecture is installable — that conversation starts here.