Notes
Writing on personal AI, building real systems, and life at the end of a dirt road. Subscribe by RSS.
-
The lanes learned to talk
The five assistants in my window can now pass work to each other: ask them all at once, let them argue, quote one into another. What that gets me is mostly honesty, and reviews I used to skip.
-
Does your agent have principles?
Three incidents from my home agent fleet — a 1.7 GB runaway file, a hardware key that proved presence but not consent, and a broken agent with a spotless audit trail — turned into three principles instead of three patches. What that chain looks like end to end, and why it's the thing worth asking any agent you run.
-
Five lanes, one window
Five AI models in one window can now hand work to each other. What that actually gets me: answers I trust more, for less money, in less time — and I still press every button.
-
The test that couldn't fail
My fleet had seven thousand lines of tests and nothing that ran them. Two models — Terra building, Opus refuting, neither trusting the other's reports — spent a morning fixing that, and the watcher caught its first real failure within the hour.
-
The payload was the judgment
Deployments used to end with a dashboard and a pager. This one ended with a conversation — a $5 Lightsail box carrying a charter-bound agent I can talk to from my phone, whose access dies with the signature that granted it.
-
Out of the loop, on purpose
The value stream maps consultants have drawn for decades just picked up a second life — a per-step test for handing work to agents, decided by cost and blast radius. Why the maps compound as models mature, and why the loops that go first are the ones nobody should want back.
-
My agent broke down. The paperwork says it never happened.
An API key expired and one of my agents quietly stopped working — while the audit trail built to record its every move said zero anomalies. A short story about the blind spot in even good monitoring, and the afternoon it took to close it.
-
I gave an agent real credentials — and a deadline
I gave an agent real credentials this week, scoped to one task with a hard time limit — then proved it live on a throwaway remote host that stopped existing the moment the job was done. What it actually looked like when the agent worked out its own access instead of me writing it, and why scoping the credential matters more than which model is running it.
-
Ask your AI what it believed last Tuesday
Every assistant now has memory. Almost none of it has a history — so when yours gets something wrong, you can't tell whether it made a mistake or was working from what it knew at the time. Here's the difference that makes, what an assistant that acts while you're asleep needs underneath it, and the three questions I use to judge whether any of it is worth your afternoon.
-
I wrote a rule against duplication, then measured 24 KB of it
One of my home-automation fleet's first principles says store facts, derive views — one canonical source of truth, everything else calculated from it. Then I measured what my agents actually load every session and found the principle had become a 24 KB example of breaking itself.
-
I hired my agent's competitors to grade its work
I build most of the ranch's software with an AI agent now. It also tells me how the work went, and I had no way to check. So I started paying its rivals a few cents to tear it apart.
-
The migration I stopped halfway, on purpose
I moved the key that signs "Craig approved this" out of a file and onto a hardware key on my desk. Then I stopped before finishing — because the safety net in my own plan was the exact problem I'd spent the day removing, and the drill meant to prove recovery works would have proved nothing.
-
The safety check I shipped, then pulled
I moved my signing key onto a hardware key, built a way to make approving things quick, and shipped a check meant to prove I'd seen what I approved. Three rounds of review later I took it back off the public repo. This is what it cost and why it was still worth doing.
-
One memory for three machines — the spec behind my agents' shared brain
Two of my agents held opposite "verified" beliefs about the same ssh route, on machines twenty feet apart. This is the spec of what I built so that can never happen silently again — an event-sourced, git-native shared memory with cryptographic operator truth.
-
Where Red came from
Our two mustangs came with federal paper trails — a freezemark on the neck, the full custody history a lookup away. Red's record and my own research notes disagreed about where he was born wild. Running down which was right turned up the story of the exact day he was captured.
-
Twelve proposals, and I never answered one
My ranch agent proposed twelve conversations in sixteen days and I never answered one — not because the ideas were bad, but because I had to context switch to answer it. Today I made answering free, and the first conversation paid for itself before 7am.
-
The day my morning brief started asking permission
I turned my morning report into a proposal feed — cards I approve by texting "yes." Day one: five approvals, one sentence that stung, and two real bugs the feed found on its own.
-
I open-sourced the layer that runs my ranch.
AI-OS Seed — the secrets vault, scheduler, and observability substrate under my ranch, extracted live from my own system and open-sourced as an alpha, MIT.
-
I built an AI chief-of-staff for myself. Here's the recipe.
One prompt, pasted into Claude Code, builds a personal AI operating system shaped to how you actually work. Free and public.