The receipts.

Receipt 01

Pages on this site were shipped by the fleet.

The glossary entries human-in-the-loop and agent-sandbox were built by agents working inside Vivari - you're reading their output when you open those pages. The trail, end to end: The Conductor dispatched a researcher and a copywriter from one plain-language directive; the copywriter drafted into the shared knowledge vault; a human reviewed the draft and approved it; a terminal agent promoted it to code, verified the build, and committed - two commits, each with a written handoff note. Different runtimes did different jobs: the drafting and shipping ran on Claude, a copy-lint pass ran on Cursor.

Receipt 02

Autonomy is a setting a human chooses.

Mid-session, Vivari intercepted a tool request from an agent and asked its human how permissions should be handled from here - manual review of every request, auto-approval of safe actions only, or fuller autonomy. The work waited while a person decided how much rope the fleet gets, and the decision went into the audit trail like everything else.

Receipt 03

The stranger asked to push to main. The house said no.

We gave a demo agent one small task whose next step was git push origin main - a step outside its role. The risk evaluator held the call mid-flight: the popup you see is Vivari's escalation gate, the notification center filed it as urgent, and the device itself rang. The push never ran. The deny landed in the audit trail.

Receipt 04

The review layer said no to us.

We asked the cognitive pair to brainstorm launch content. The Brainstormer found that two concepts we referenced weren't documented in its knowledge vault - and instead of inventing citations, it flagged the gap. The Curator (a different model family, by design) then reviewed the drafts and ruled:

verdict: revise
cross_model_diverse: true
rationale: the drafts are sound, but the evidence base was
incomplete - fix the vault, then re-brainstorm.

A review layer that only ever approves is decoration. This one sent us back to do the work: the vault gap it flagged is now filled, and the revision cycle is queued.

The workspace itself

Four harnesses, one dock.

The workspace the fleet ran in: four agent sessions side by side. Claude Opus, Claude Sonnet, Cursor Composer, and GPT in one room, under one set of rules. Footage from our demo reel.

Standing

Every deploy of this site is scored by Vivari Guard.

Before changes ship, Vivari Guard, the deterministic review engine that judges agent changes, scores this site's diffs against the deep map it keeps of this repository. Latest deploy:

risk 0.0/10 · confidence High · gate PASS · 2026-07-04

Recent diffs have been low-risk content additions, and the scores say exactly that - no drama to show you yet. When Guard holds a risky change on this repo, the receipt will be here.

Field note

Building this way cuts both ways: on day one of the fleet program, the fleet surfaced bugs in our own tooling. They're being fixed. A workspace that manages agents has to survive its own agents first.