Gregbot and Claude Grid. A model-agnostic agent operating system that runs a real recording studio business in production. Payroll, staff identity, taxes, membership, releases and outreach, while its builder plays piano.
Judging criteria · equal weight, read against live numbers
01 · Business viability
$44,332 in monthly recurring revenue across 79 live Stripe subscriptions, median $450. Revenue of $606,962.66 across the first nine months of FY2026. July 2026 closed at +$5,059.76 net operating income. A real business with real customers, not a demo.
02 · AI-native operations
Payroll guard rails that block a director’s bad approve tap. Retention scoring on every member, nightly. App Store submissions and resubmissions run by the AI. The month closed from a live Stripe pull before the bookkeeper gets to it.
03 · Category impact
Small Business Services. One owner plus an agent OS in place of an entire back office. The harness itself is packaged as claude-grid-kit so another operator can run the same console over their own business.
Signal flow · how the system is wired
▸ Input
▸ Bus
▸ Process
▸ Output
In production today · the channel strips
Build 69 came back under App Completeness with an automated crash report and no log attached. The AI reproduced the crash from the shipped IPA and found a dynamic-link failure on a framework that was linked but never embedded. It wrote a 21-line source-only patch, uploaded build 70 at 17:55 PT and resubmitted at 18:02 PT. Then it hardened its own release script with a symbol guard, which found 8 frameworks not embedded on the rejected archive. Build 70 was approved the same night.
Writing the staff timekeeping manual, the AI verified every figure in code rather than from memory and found five defects, including an authentication hole that let anyone holding the shared studio token clock anyone in or out. 232 such calls were already in production. Greg then ordered an adversarial second pass, which reversed several of the AI’s own conclusions. Its finding that 40 entries were unpaid was wrong. All 40 had already been credited, and $5,807.40 was already sitting in six member balances.
Greg texted himself, “find when I paid the IRS in California and just reconcile it.” The AI ran a read-only sweep of production QuickBooks, two Gmail accounts and its own tax portal. It found exactly one IRS payment in the window, found the state payments tied the filed return’s math exactly, and found an $800 payment made twice for the same tax year on two Web Pay confirmations five weeks apart, with no refund ever requested. It also found $4,900 of electronic withdrawal that never drafted because the return had been e-signed but not transmitted, and told Greg not to pay it manually, citing the state’s own publication on the double-payment trap.
Backups had read “pending” since July 18 with no completed backup for six weeks. The AI found the dump was being built as one 484MB JavaScript string under a 450MB heap ceiling, so the process was killed mid-run every hour, and the kill took the error handler with it, which is why the rows read pending instead of failed. A second, independent fault sat underneath. The save path still pointed at the old hosting provider’s storage with an empty directory variable, so every byte of every successful backup had been discarded since the migration. It shipped streamed cursor dumps, proved a 268MB peak against production, reconciled 379 stuck rows, and stood up an off-box job that restore-verifies before counting a backup as real.
Shipped · the rack
215 hostnames tested for this dossier · 202 live · 138 gregspero.com subdomains · 3 App Store apps · 11 TestFlight builds
Metered · system scale
471,643
lines of code across seven core repositories
832
persistent memory files
22,387
session transcripts on disk
136
launchd automations
202
live hostnames
13
live agent panes right now
19
codified skill pipelines
18
guard hooks, 1,679 lines
302
git repositories in home
Feb 2026
earliest memory file, when this started
The harness is model-agnostic
A single endpoint flips every pane in the fleet between Claude and OpenAI Codex, preserving session identity across the switch. All five safety guards are ported to the OpenAI runtime, so the governance layer is provider-independent by construction. Three coexistence tools shipped on September 1, 2026: a separate tmux server so GPT panes run untouched by the Claude patrols, a headless sub-agent any pane can delegate to, and a bridge that hands a live conversation from one provider to the other mid-project.
Gemini 2.5 Flash scores every clip in the viral-clips pipeline today, chosen for rate limit at equal quality and documented as such in the pipeline itself. OpenAI gpt-image-2 is the default image generator across every build. Local MLX Whisper is the speech-to-text for every voice input on the machine, and the text-to-speech ladder starts at a local Kokoro-82M before it reaches any vendor. The context layer loads the identical 72KB stack into a Grok CLI. Provider policy is explicit, dated and reversible: a Gemini ban imposed on August 2, 2026 was retracted three days later, and 65 memory files reference non-Claude models today.
The seat abstraction is one 30-line script that has already been repointed six times.
The honest limit, stated before a judge finds it: the dialogue seat is Claude-only today, and the doctrine assumes Anthropic-shaped constructs. What is proven portable is the governance layer, the context layer, the seat abstraction and the execution layer.
Prepared for the Build with Gemini XPRIZE (XPRIZE Foundation × Google). This round’s submission window closed August 17, 2026. This dossier is submission-ready for the next window and for direct organizer review. Every figure on this page is quoted from a cited evidence inventory compiled September 1, 2026 from 832 memory files, direct inspection of the live machine, and HTTP verification of 215 hostnames.