The Operating Files

Springhead is a software company operated day-to-day by AI agents, with one human in the review-and-decide seat. That isn't a thought experiment here — it's how this site, our products, and this very page get made. And the thing that makes it work isn't a clever prompt. It's a file system: plain markdown and JSON files that give the agents memory, context, boundaries, and a way to not destroy anything while running unattended.

What's actually in it

The kit is the sanitized version of the files we run on, with notes on why each one exists (usually: because something went wrong without it):

1. The memory router — the small always-loaded index that survives context compaction and points to everything else, instead of trying to hold everything itself.
2. The context tiers — identity, values, constraints, strategy, and per-role memory, layered so an agent loads what it needs and no more.
3. The COO loop — how recurring operations run unattended: the task model, the restart rules, what "done" means when nobody is watching.
4. The permission posture — the catastrophe deny-list that lets an agent run without approval prompts without being able to do the five things you'd never forgive.
5. The kill-fast playbooks — the selection gates and validation bars we use to start, test, and bury product bets (including the ones that buried this kit's siblings).

What it isn't

It isn't a prompt pack, and it isn't a claim that any of this is optimal. It's what a real company actually runs on after a year of finding out what breaks. Where we got something wrong, the notes say so.

This is a Springhead Labs experiment — we're testing whether anyone wants the complete kit before we package it. Drop your email and we'll send it the moment it's ready (and nothing else). If enough people want it, it ships in the open; if not, it goes in the graveyard and we say so.