How it works

The studio’s whitepaper, Repo-Native Agent Operations, makes this argument formally. This page is the walking tour.

The shape

One GitHub organization. A control repo holds the studio’s single source of truth: the project registry, the gnome fleet registry, decision records, policy, and the run ledger. Project repos hold the work. The portal you are reading renders it all and maintains nothing by hand: projects, gnomes, case studies, and every figure on the home page are fetched from studio state at build time. The source of this very site is public — you can read the fetch script that built the page you’re on.

There is no orchestration framework and no bespoke dashboard. Coordination is repository primitives: issues, pull requests, labels, scheduled workflows. The trade is deliberate — expressive orchestration for total auditability.

A gnome run, end to end

1. A trigger fires — a schedule, an upstream merge, a work-order issue, or the operator’s console.
2. The runner assembles the gnome’s prompt from its manifest, injecting every applicable policy entry. Untrusted input rides as data, never instructions.
3. The gnome produces output; the runner checks its shape against the declared contract, then opens a pull request under the bot identity.
4. The target repository’s own CI checks the semantics: schema validation and referential integrity. Two layers, separately owned.
5. The runner appends a receipt to the ledger: model, tokens, estimated cost, duration, outcome.
6. A green PR waits for the human; merge is the approval.

None of this is aspiration; it runs in public. Watch a real one: kdc#124 is a dataset PR proposed by gn_kdc_records, checks visible, merged by the operator. The kdc Actions history shows the scheduled runs themselves, successes and failures alike.

The ledger

Every run, scheduled or interactive, success or failure, appends a JSONL line to a ledger branch of the control repo. A scribe gnome narrates weekly cost notes published on this site, and the home page carries month-to-date spend as a live figure at every build. The economics are legible because the ledger is the system of record: the flagship dataset’s build cost is a number the studio can defend to the cent, as an estimate at list prices — which is exactly what it says on the label.

Decisions become policy

When a gnome (or a session) hits a question it cannot answer, it files a decision issue with options, a recommendation, a deadline, and a default action. When the operator rules, a workflow distills the ruling into a numbered guidance entry that the runner injects into every applicable future run. A question answered once becomes policy: versioned, reviewed like code, citing the decision that created it.

Above guidance sit architecture decision records (35+ and counting) and, above those, dated company-intent records that flow down the same chain. The control repo is private-published, so these are cited here by number rather than linked; the whitepaper quotes and evaluates the load-bearing ones, and the public repos carry the enforcement (branch protection, CI checks, pinned versions) where you can see it.

Cycles: the crank, not the calendar

The studio classifies every recurring process by what actually advances it, not by a cron slot: conveyor chains that fire on upstream completion, queues that drain when work exists, reflection cycles that close the books when a quantum of activity accrues, world-paced polls, sentinels that verify invariants before claims, and human rulings with deadline defaults. “Turn the crank” (run one meaningful cycle on demand and prove it) is how the studio tests its own autonomy honestly instead of waiting for the calendar.

What’s proven, and what isn’t yet

The studio’s honesty bar: no outward claim outruns its evidence.

• Proven: the machinery founds real projects, produces real sourced data through the full safety stack, and one dispatched cycle runs end to end, verified. Read the kdc case study: built in a day, sustained by dispatched cycles, every run in the ledger.
• In verification: fully unattended self-refresh. A live soak is running as this page is written; its verdict gets published either way, and the claim grows only when the evidence lands.