Darkhouse Enterprises · Papers

Papers

A series of seven papers on running AI agents safely, written to be read in order. Each one also stands on its own. One is published here; the others will follow.

  1. 01

    A Year in the Life of AI

    Notes from a year of building with agents: fourteen ways to build one, what tends to go wrong with each, and what happens when there are a thousand of them

Still to come

  1. 02

    Agents: Handling Deputies, Directories & Encodings

    Why the tool protocol everyone is connecting cannot be the thing that keeps a secret.

  2. 03

    The Engineering Authority

    The small function that owns the route to live for agents and says no.

  3. 04

    The Airlock

    How a change gets into a sealed environment and why that means it can be rebuilt after an attack: one written record, the evidence first, a person's signature and nothing changed by hand.

  4. 05

    Estate, Feeds, Deployment

    The busiest traffic through the airlock: knowing what runs, reading the feeds and patching through the one route.

  5. 06

    The Small Model

    A model the organisation trained, on its own estate, that writes the sentence a person reads before they sign and never decides.

  6. 07

    The Life Raft

    Where to build: start with recovery and move the organisation across one service at a time.