Build Logs

I gave my AI agent a script and eight rounds of notes. It made a nature documentary.

The Wild Agents is a spoof nature documentary: a small robot in a misty forest gathers forty-seven "dependencies" until a beetle-sized predator knocks the pile over. It was built in Blender, a free 3D program, by Earl, the AI agent that runs my lab, from my script and my notes. I've never built anything in 3D or edited a video. This build log is how it got made anyway: the words first, a borrowed cast, gray before color, my notes, the rendering bill, and twelve versions, each a little less wrong than the last.

My AI agent built a music video in 43 hours. Every version taught it something.

I had an idea for a club track that laughs at empty showing off. Forty-three hours after I handed my agent, Earl, the brief, it was a finished concert film: a song made with Suno, an AI music generator, more than 6,000 stick figures dancing to recorded motion capture, a stage built in Blender, a free animation program, and twelve numbered versions in between. Three more versions followed the next day. Every tool in it came in because the version before it showed a problem. This build log is that sequence, frame by frame, including what broke.

My AI agent fell asleep every time I closed my laptop. So I moved it into a Mac mini.

Four companies shipped an always-on AI agent in ten weeks, and all four run it on a computer in their own cloud. That's convenient, and it costs reach: an agent in a data center can't touch the servers, file shares and tools in your house. On Oct 22, 2026, I moved my own agent, Earl, onto a Mac mini at home instead. A week later I moved half of it back. Earl now works from two machines: the laptop when I'm at the desk, the Mini when I walk away. This build log is how that works, what it cost, and what I'd want to know before setting it up. It isn't for everyone.

My AI agent couldn't see what I meant. So I gave it a screen that disappears.

My video skill had no screen, so every note I gave it was a sentence. Sentences are bad at pointing. Instead of building an editor, I built a page the agent opens beside the chat only when pointing beats talking, then puts away. Here's the idea, what building it taught, and why it reaches past video.

This video? Same agent.

The Lab needed a reel that says what it is and why it exists. An agent made all of it: the script, the music, every diagram and title, the recordings of the site, three cinematic shots, the edit on the beat, and the checks before anyone saw it. I approved and gave notes; I never touched a timeline. This entry is how that works, what twenty-one rounds of review taught, and a playbook with the skill that lets your agent do the same. The whole reel cost $12.84 in generation.

The free model at the front of the line

A decision model small enough to run on a Mac mini, with open weights and no bill, can in principle take the first look at every request and pass only its doubts to a paid model. This entry is about when that cascade pays: what the free tier has to get right, how to tell whether its confidence can be trusted, and what one two-hour fine-tuning run on free GPUs changes. The evidence is three experiments on my own hardware against the paid model from Build Log No. 005 and a frontier model, for $5.45 in total API spend.

The fastest-adopted AI model can't write a sentence. Here's what it's for.

In its first day on one of the largest model gateways, a model that can't produce a word of prose was adopted faster than any model before it. It answers only the questions you write in advance, with odds attached. This build log is about why that combination (messy text in, a bounded yes, pick or rating out, for a fraction of a cent) may be a new building block for ordinary software, and what it does to the cost of deciding anything by machine. A working demo and two dozen measured runs on my own hardware are the evidence. They earn the argument where they can and mark its limits where they can't.

The work is not a line

Every delivery process I've ever worked in is shaped like a line — plan, design, build, test, release, operate — and most plans for adopting agents just slot them into it. So I put a meter on three days of real work with my agent: 185 activities, every one classified and plotted. Then I had a second model attack the dataset and a third blind-rate a sample, so what's here survived cross-examination. If you're deciding how agents fit into your ways of working, this is the shape of the thing you're planning for — and it is not a line.

The work before the work

A project that hasn't been built yet, written up anyway. The order I plan in, why the design has to cover running the thing and not just standing it up, and what happens when I hold my own method up against the findings on this site. Four of my habits survive it. Four don't.

Agent observability and performance

Past entries keep saying there's an agent doing the work around here. That's been talk. This is the instrument panel I actually use to watch it: what it costs, what it remembers, what it can do, and how it gets tuned. Real screenshots, sensitive bits blurred, everything else as I see it.

Hands last

The control panel was the goal from day one, and it was the last thing built. Three weeks of read-only discovery came first, on purpose. The machines being wired up were the ones that couldn't be replaced.