Building Loomfield: an operating system for AI work
Why the next useful layer for AI is not a better chat window, but a shared brain, durable work, and controls that earn trust over time.
Series
The Loomfield build log
Loomfield is a self-hosted operating system for AI work that I'm building. Every two weeks I write down what changed, what broke, and what I learned. This page collects every entry in order.
What is Loomfield?
A shared brain, a project model, and controlled gateways so different AI agents can work from the same durable context. Read Building Loomfield for the full idea, or visit loomfield.ai.
Start here
Small Context, Serious Work
How Loomfield uses focused retrieval, smart micro-sessions, and the smallest capable model to spend context where it actually improves the work.
Preparing Loomfield for the Public
What changes when a private-beta AI work system has to become clear, trustworthy, and useful without its builder in the room.
Build log entries, newest first
Build Log: Workflows and a Hub That Kept Freezing
Visual workflows for building automations without code, and a two-week hunt for the reasons the hub kept freezing as it grew.
Build Log: Loomfield Leaves the Room
The first invite-only alpha builds for Mac and Windows, the install bugs my own tests missed, and a rebuild of how decisions are tracked.
Build Log: Stop Shipping Dark
Two weeks teaching Loomfield to learn reusable tools from its own work, fixing a status bug properly, and ending the habit of shipping everything switched off.
Build Log: The Audit That Audited Itself
A reusable audit for checking whether AI-finished work is really finished, the bugs it found in itself, and cutting the cost of re-sending context.
Build Log: Making the System Keep Its Own Books
Two weeks turning record-keeping from a rule agents follow into something the system checks, plus much faster recall and a hard look at routines.
Build Log: One Version of the Truth
Three weeks moving Loomfield's memory onto structured records, rebuilding Mission Control around honest screens, and having more than one model check the work.