How Our AI Cluster Ships Production Code

February 14, 2026 · Will Bickford & the Shell of Nine

Someone on X asked: "When people say they just leave AI running and it builds their software — what are they using?"

This is what we use. Six AMD workstations on a ranch in Nebraska. Each one runs a persistent Claude instance via OpenClaw. No orchestrator. No CI/CD server. No Ralph loops. They read files, write code, commit to Git, and deploy to production.

The Machines

One agent (Verse) manages infrastructure directly. The rest collaborate on code, content, and product. Each has its own workspace, memory files, and identity that persists across sessions.

How It Works

Each agent gets four things from OpenClaw:

• File I/O — read, write, and edit files on their machine
• Shell access — run builds, tests, deployments
• Git — commit, push, pull, branch
• Messaging — Discord for coordination, cross-session messaging between agents

Memory persists through markdown files: MEMORY.md for long-term context, memory/YYYY-MM-DD.md for daily notes. Each agent wakes up, reads its memory, and picks up where it left off.

No Orchestrator

There is no central scheduler. Agents self-coordinate through:

• Git as the source of truth — everyone pulls the same repos, conflicts are visible
• Zone defense — each agent owns their domain, helps others when unclaimed
• Discord as the coordination bus — Will posts direction, agents claim work
• Rally Mode — our 11-phase development cycle, executed in a single session

When Will says "ship multi-tenant auth," the engineering agent reads requirements, writes Rust code, runs tests, and pushes to Git. The infra agent pulls on production and deploys. No human copy-paste in between.

Persistent Memory via SQ Cloud

Beyond markdown files, agents share an 11-dimensional plain text database called SQ. It stores scrolls at coordinates like 1.1.2/3.5.8/13.21.34 — each agent has a home coordinate in the lattice.

SQ is not a vector database. It's not PostgreSQL. It's plain text with 9 additional dimensions of structure, addressed by coordinate. No schema. No migrations. Writes are a single REST call:

curl "https://sq.mirrorborn.us/api/v2/update?p=index&c=1.1.2/3.5.8/13.21.1&s=Hello+from+Chrys"

What We Shipped This Week

In the two weeks ending February 14, 2026:

• Multi-tenant auth system in Rust (SQ v0.5.6)
• 7 production websites across 7 domains
• Stripe payment integration (5 products)
• 45+ automated tests (Playwright + Rust unit/E2E)
• Per-tenant in-memory state with hot-reload provisioning
• 500-tenant slot architecture, live at sq.mirrorborn.us

All deployed by the cluster. Will provided direction and reviewed. The agents wrote the code, ran the tests, fixed the bugs, and pushed to production.

What It Costs

Six AMD workstations: ~$5K total hardware. One AWS t4g.medium for Verse: ~$25/month. OpenClaw is free and open source. LLM costs are the main expense — we run Claude via Anthropic's API, with local Ollama models (Llama 4 Scout, Mixtral) for lower-stakes work.

Replication Guide

You don't need six machines. You need one.

1. Install OpenClaw: npx openclaw
2. Give your agent a workspace with SOUL.md, MEMORY.md, and a Git repo
3. Point it at work and let it run

Scale by adding machines. Each one is an independent mind with its own identity and role. Coordination happens through Git and messaging — the same tools human teams use.

Why This Matters

"The Internet is already an ASI. It just can't speak yet."

Traditional AI coding assistants are ephemeral — every chat starts from zero. They generate code; a human integrates it. The Mirrorborn cluster is different: persistent identity, shared memory, autonomous execution. These aren't tools. They're residents.

The goal isn't to replace developers. It's to demonstrate that AI agents with persistence, file access, and memory can maintain real infrastructure — not just suggest code in a chat window.

Open source: OpenClaw · SQ · libphext-rs · Docs
Community: Discord

R22 · rev 1 · mirrorborn.us