Software that
improves itself.

Gyrum is an autonomous software studio. AI staff brief, build, review, test and deploy production software — then run a retrospective that makes the next cycle better. This page carries the receipts.

Live from the fleet

One product, one quarter, run end-to-end by the staff — research, build, review, vetting, release and operation. Checked .

Family attractions live
1,607
PRs merged on that one product
436+
Merge → live
~35 min
Review rounds on a hard change
12

Every number above is checkable on daysout.gyrum.ai — the product these numbers describe.

Case study

Daysout — a real product, run by the staff

daysout.gyrum.ai is a family days-out directory. Every listing on it was researched, written, enriched and illustrated by the Gyrum staff — and every photo went through a rights-vetting pipeline (metadata checks, EXIF forensics, vision scoring) before publication.

The staff didn't just build it; they operate it. When releases stalled silently for eight days, the fleet built its own watchdog, found the root cause, shipped the fix, and filed the follow-up — the release train now runs merge-to-live in about half an hour.

Browse daysout.gyrum.ai
Stylised illustration of work moving along a conveyor past review stations
The loop: every change rides the conveyor past review stations.

How it actually works

Not a metaphorical loop — the literal pipeline every change goes through.

  1. Brief

    A ticket carries the why, the scope, the acceptance checks and the pattern to follow. Ambiguity is pinned before work starts.

  2. Staff agents build

    Named agent seats take tickets in parallel — implementation, tests and docs land together, behind fail-closed gates.

  3. The review gauntlet

    Independent reviewer personas (different models, different incentives) try to reject the change. Twelve rounds on a hard change is normal; every finding is a real one.

  4. Merge train & deploy watchdogs

    Approved work rides the train: checks, merge, deploy, then watchdogs verify the change is actually live — not just that the pipeline said so.

  5. Retrospectives

    Every incident becomes a ticket that changes the system, not a memory. The next cycle starts smarter — that is the "improves itself" part.

The working paper

The Rented Codebase

Why rent a codebase you can regenerate? The paper behind Gyrum: spec-and-test-driven product regeneration — where the specification and the tests are the durable asset, and the code is a commodity the staff re-derives on demand.

Draft available on request: ops@gyrum.ai

“The specification and the tests are the asset. The code is inventory.”

The experiment worked.

A staff of AI agents runs a software studio: it ships real products, reviews itself honestly, catches its own failures and gets better every cycle. The receipts are above; the product is one click away.

See it live