The 408A house that runs on AI

Field report · · The 408

Leave the porch light on: opening 200 field reports from a house that runs on AI

A porch light on the record: more than two hundred field reports, each with what was planned, what an agent claimed and what the house showed.

Behind The 408, a real house run by an assistant called Fred, sit more than two hundred field reports. Each one records what happened when AI agents worked on a place where people live: what was planned, what an agent claimed, and what the house actually showed. If you build with AI agents, they’re the most useful thing we have, and until now nobody outside the house had read any of them.

A porch light says someone’s home and you’re welcome to come up the steps. This is us turning it on.

200+field reports from real builds on a real house
3rules that run everything else
7 daysminimum before any incident is written up, and only once it’s fixed
0addresses, schedules or security details, ever

Why bother

Demos end when the video does. A house doesn’t.

Nearly all writing about AI agents comes from demos and benchmarks. Here, a wrong change turns off someone’s lights, and the people who live there never agreed to be the test team. That makes the house an unforgiving test bench, and an honest one: the light either changed or it didn’t.

A real houseLights, heat, locks and the people who live with them. Every mistake lands on someone.
200+ field reportsWhat was planned, what an agent claimed, what the house actually showed. Failures included.
A field manualFor supervising AI in the place people live, written from the record instead of a demo.

What this isn’t: a benchmark. It’s one house and the record of what broke there.

Earned one bad afternoon at a time

Three rules that run the house

  • R1

    Don’t treat a report as evidence of success

    An agent’s “done” is a claim. The house confirms the change, or it didn’t happen. Once that rule existed, everything else got built around it: checks that read the real state back, gates before anything ships, and reviewers who weren’t the author.

  • R2

    One writer per machine

    Two agents once worked on the same server at the same time, and a few hundred movies dropped out of the media library. They were back in about ten minutes. The rule it left is simple: one agent changes a given machine at a time, and everyone else reads.

  • R3

    Nobody grades their own work

    The agent that built something doesn’t get to say it works. A separate check does, and for anything that touches the house, a person does.

Come up the steps

What you’ll find here

  1. The build log. What went wrong, why, and the rule it left behind. Every post is written up after the fix, never live.
  2. How Fred works. The design behind an assistant that asks before it changes anything, can undo what it did and keeps a journal. Fred’s brain is a good place to start.
  3. Porchlight Starter. A free kit that makes an AI coding agent careful. Coming soon.

What stays inside

Where the house is, when it’s empty, how it’s secured, and anyone who lives there. The light is on for the work, not the family.

The light’s on. Come up and look around.

One question to take with you: what’s the one rule your agents follow because something went wrong?

← All field reports