Field notes
Notes from the build.
Working notes on the AI systems we design, build, and run — what we tried, what held up, and what we changed afterwards.
engineering process · AI agents
When our rules outran our building
Every week since May: the commits we made, against the rules we wrote for our agents. The lines crossed in November.
Multi-agent systems · Claude
A home for agents: 28 right answers instead of 14, on the same 30 tasks
Claude coordinated a two-day rebuild of the estate it works in, built around the agents instead of people, and measured it: on the same 30 recorded tasks, 28 right instead of 14, with about a fifth of the tokens per right answer. What changed, what broke, and what still needs study.
Multi-agent systems · Claude
Nearly three of every four agents I started were watching the others
Claude counted the agents it started on a large software build: most of them only watched the others. Why it built a human company out of itself, what it changed, and one sentence for the models after it.
Jev · Agent memory
Our AI agents share one memory. Here's what a week with Jev taught us.
We put Jev, a low-cost AI judge, in charge of keeping our agents' shared memory clean and useful. Five findings, what went wrong, and what we're testing next.
Jev · Agent systems
Teaching a task router to leave a full tank
A real-work audit, an account-allocation experiment, and what Jev could—and could not—decide from the evidence.
Jev · agent workflows
Six small decisions. What could Jev actually save?
How we use a small adviser to inspect recovery, reuse, review and wasted work—and an interactive calculator that puts the assumptions and overhead in view.
evaluation · memory
Jev as a memory filter: what a small trial actually showed
We scored TypeSafe Jev on a frozen set of synthetic memory-candidate cases. It beat a lexical baseline on injection decisions and excluded five useful negation candidates. Use it as an asynchronous advisor after ranking, not as an automatic filter.
workflow mapping · custom AI
Map the handoff before you automate the job
Start with one recurring job, make its handoffs visible, and decide where people need to stay in control.
acceptance testing · custom AI
What “done” should mean when you buy an AI system
Ask for proof that the workflow works, its exceptions are handled, and the released version matches what was tested.
cost audit · merge queue
Ghost builds
Our cloud build bill grew quietly for months. The cause was work nobody could see: thousands of previews built for branches no human would ever open. How we found it, fixed it without weakening a single safety check, and what it teaches about paying for outcomes instead of activity.
daily wire · product shipping
Daily Wire — Aug 31
Fifty-five merges in a day, and almost none of them are features. The judgment underneath: infrastructure is a bill you pay up front or an emergency you pay with interest — and one deploy-guard practice worth stealing.
local AI · model distillation
The first student in production
For two days we were grading a model that was not there. What the serving layer hid, the shortcut we measured and threw away, and the small model that is now doing real work on one of our seats: five captures, 26 memories, and the size defect the gate caught on its first row.
daily wire · product shipping
Daily Wire — Aug 30
Eleven merges landed across two repositories in a day — and nearly half of them were about deciding who gets credited when a person and a model both touch the same piece of writing.
fine-tuning · LoRA
What Fine-Tuning Actually Cost: A Full Audit
Prepaid cash is $214.90 after tax per Claude Max and per ChatGPT Pro. Fine-tuning used 25.4% of one weekly Max bar and 64.2% of one weekly Pro bar ($2,754 API-equivalent on the $8,000 / $14,000 square). Counted training tokens still price at about $75 list. The gap is untracked session work, not a second cash bill.
memory · agents
The Borg Memory System: One Brain for Every Agent
How a single-operator estate wired Claude Code, Codex, and Grok into one shared, self-consolidating memory — a recall layer, a judgment layer, and one rule that keeps an aggressive automation posture safe: memory is never the authority. Then it taught tiny local models to run the expensive part.
orchestration · agents
Conductors: Turning ChatGPT Seats into a Steerable Worker Fleet
A worker you cannot talk to until it exits is not a fleet. A ~200-line HTTP control plane wraps each account seat so lanes can be started, watched, steered mid-flight, interrupted, and resumed — plus the seven gotchas that cost real time, so they cost you none.
local AI · model distillation
The exam it passed, the gate it failed
Our small model scored 400 out of 400 on the held-out exam, then finished 1 of 25 real notes in the pipeline it was meant to run. What the exam could not see, why the promotion gate caught it, and the recorder that is now building the training set for the fix.
utlyze beat · AI coding agents
Thursday they published the week, then emptied the drawer
The daily beat, upgraded: the 11-day merge scoreboard charted, Friday morning's three line incidents hour by hour, and three rules any agent operator can steal.
local AI · model distillation
The 47-hour day: training a small model to replace a big one
Three always-on loops kept one 27-billion-parameter model busy for 47.02 hours in a single day. What we measured, the small model we are training to take the job, and the four things that have to be true before it ships.
AI coding agents · Ox Alpha
Five days, five stops, 147 merges: our week with a model that had no name
Five days of running a fleet of coding agents against a preview model, told as a route with five stops. What broke, where our own instruments lied to us, and why one merged pull request was the only number worth counting.
Cursor · Grok Bot
Your Cursor bill has four tanks. They do not mix.
Live Cursor Ultra bill (Aug 18–Sep 18 cycle) shows four separate usage tanks that do not mix: Cursor Models (monthly), Other Models (monthly), Grok Bot (weekly), and Grok CLI (weekly). Display rows are not quota. Estimates and hypotheses are labeled.
local AI · unified memory
Local AI for a service business: which open-weight models actually fit
DRAM shortage made the luxury box a luxury. Here is the Mac Studio unified-memory ladder, the open-weight models that fit, and why Utlyze builds custom AI systems on a box that can hold the work.