The Operating Files

You've worked out that AI can do real work in your business. The hard part is the part nobody sells you: turning "I ask an AI to do things" into something that runs — that remembers across sessions, follows your rules, keeps working while you sleep, and doesn't wreck anything when no one's watching.

That's an operating system, and we built one to run a whole company on AI agents — this site, the products, the day-to-day. This is that system, and you can put it to work on yours.

What it actually does for you

Four things that turn an AI from a clever assistant into an operation:

1. It remembers. Your agents stop starting from zero every session and stop drifting when the context fills up — a memory that persists and a place for everything they need to know.
2. It runs unattended — safely. A loop that advances work on a schedule and reports back, plus a guardrail setup that lets it run without you hovering and without the 3am disaster.
3. Your judgment compounds. The calls you make over and over get written down once and applied the same way every time, instead of re-decided in every session.
4. It's yours. No platform to be locked into — the whole thing is legible files you can read, change, and keep.

Every part earns its place by fixing a specific way unattended agents fail — and ships with a plain note on why it exists, usually because something broke without it. That's the part you can't shortcut: not the files, the scars in them.

A piece of it, free: the never-do list

Here's the file people ask about first, in its public form. Our agents run unattended in bypass-permissions mode — no approval prompts — paired with a short deny-list that's enforced even in bypass. The design question for every entry: what would we regret at 3am?

One honesty note before the list: publishing your real deny-list is reconnaissance for an attacker — the value isn't the entries, it's the gaps (exact knowledge of what isn't denied). So this is the categorized, illustrative version — deliberately not our live config — and that lesson is itself part of the system.

• Destructive file operations — recursive/force deletes outside scratch dirs
• History rewrites & force pushes — anything that can destroy the audit trail
• Privileged execution — sudo, system installs, service restarts
• Secret reads & exfil shapes — credential-file reads, secrets piped into network calls
• Constitution-path writes — the agent's own rules, personas, and permission config are human-writable only
• Money & outward actions — anything that spends, posts, or emails, unless that exact action is the sanctioned job and separately gated

Keep it one screen long — a deny-list with forty entries is an allowlist in denial. Every line of ours traces to a real scar (one of them is the day a "cleanup" tool ran on live data; it did exactly its job — that was the problem). That's the texture of the whole thing.

Autoresearch, reworked for business questions

In March 2026, Andrej Karpathy open-sourced autoresearch — an AI agent that runs hundreds of experiments overnight, keeping whatever improves a number and discarding the rest. It's brilliant for anything you can score: training loss, build times, render speed.

But most business decisions don't have a number. "Should we add TikTok Shop?" doesn't hill-climb. So we kept the engine — set a direction, let agents explore in parallel, keep what holds up — and rebuilt it for the calls that turn on judgment instead of a metric.

Karpathy's version optimizes. Ours explores. And since there's no score to chase, your CEO sits with you first — sharpening the question and the rubric — before the research runs.

Autoresearch for the decisions you can't score.

Two ways to put it to work

Run it yourself — $49. The complete system as files: memory, context, the loop and its dispatch rules, the guardrails, the playbook layer, and an installer that scaffolds it into your own repo in one command. One-time, downloaded, yours to adapt and keep. No subscription, no lock-in.

Get the files — $49 →

Or have us run it for you. Don't want to wire up files and babysit a loop? Skip the assembly: your operation provisioned and run for you — the memory, the loop, the dashboard, the guardrails enforced live — so AI is doing the work, not waiting on you to prompt it. We run exactly this for ourselves every day; we haven't opened it to anyone else yet. We'll build the hosted version if enough people want it, and say so publicly if they don't. Drop your email if that's the one you want — it's all the email is for.