I run twenty projects with AI agents. My job is to be needed as little as possible.
The agents write the code, run the checks and post the alerts. My work is to decide what they may do, catch them when they drift, and keep my own part of the job small. This page shows how, with real numbers from my own records.
Every job I have scheduled, on one chart
Select any point to read its registry entry. Larger points run more often.
- job
- loop-health
- area
- Operations
- status
- live
- runs
- weekly
- engine
- Claude Haiku
- check
- yes: a command that fails if the job did not do its work
- data leaves to
- 1 named service
Each point is one entry in my job registry, as it stood on October 18, 2026. Retired jobs stay on the chart, because the record of what I stopped is as useful as the record of what runs. The grey points are private jobs; their details stay off this page. Storefront and content jobs show only their area and schedule.
“By August, there were six times more agents than employees at Anthropic, and productivity increased.” “All this acceleration means that, increasingly, the bottleneck is human attention.”
I hit the same limit with one person and a laptop. Agents made new work almost free, so the work grew until checking it took all my time. Four problems came up again and again. Here is each one, what I built to fix it, and what it has done so far.
Jobs piled up faster than I could explain them
The problem
An agent can schedule a new job in a minute, so I let them. In ten days the count went from 13 to 39 to 53, and I had retired none. I could no longer say what half of them were for, or what would break if one stopped.
What I built
One registry of every job, and a cap on how many may run. I cannot raise the cap without writing down why.
How it works
- Each job gets an entry: its goal, the check that proves it ran, the model it may use, its spending limit, and where its data may go.
- A check fails if any job runs without an entry.
- A file holds the cap. Raising it means writing the reason in that file, where the next review will read it.
- A weekly job runs every job's check and reports the ones that fail.
I found three cloud jobs running that I had forgotten, and put them on the record. Today 66 run, exactly at the cap.
A team with six agents per person needs this list before it needs more agents.
An agent's instructions can change and nobody notices
The problem
In August a developer asked an assistant for an app, pasted a one-line install command from a copycat site, and the malware wrote a poisoned skill file into their agent setup. Every later session obeyed it. My own machine showed how easy that would be to miss: one weekly sweep found 2,148 new files in the agents’ folders. All but 25 were leftover plugin files and harmless, but I only knew that after reading them.
What I built
agent-integrity, an open-source tool that records every file an agent reads its instructions from, reports any change, and asks me to sign only the ones that raise risk.
How it works
- It takes a fingerprint of every file that tells an agent what to do: each skill, hook, settings file and tool server.
- Each check compares today's files with the last signed set and explains every change in one plain line.
- Changes are sorted by risk. A new permission or a new hook needs my signature. A theme change settles on its own after a day.
- A local model may write the explanation, but it can only raise a risk, never lower one. The file it reads could be the attack.
- Version 0.4 puts the same signature on scheduled jobs. If a job's script changed after I approved it, the job does not run.
- It once blocked three jobs that touch money, hour after hour, while the weekly health check still said green. A refusal should never look healthy, so I have written a change that turns the health check red; it is waiting for my approval.
It watches 15,096 files on my machine. Its 172 tests pass.
When agents outnumber staff, nobody reads every instruction file. Something has to know which changes need a person.
“It builds” is not the same as “it is ready”
The problem
Agents built my nine sites at different times from different prompts. A site could pass its build and still have no privacy page, no sitemap, or a key left in its shipped code. A clean build told me nothing about any of that.
What I built
A 20-item launch standard, and a script that checks it mainly against the live site, not the code on my laptop.
How it works
- The script loads the real site and samples its pages.
- On most sites, 17 items are checked by machine: privacy and terms pages, HTTPS, secrets in the shipped code, sitemap, alt text, broken links, spam protection and the rest.
- Speed, contrast and phone layout need a separate browser audit. Until it runs, they are marked open, not passed.
- Any red item fails the script. A red item means the site is not done.
- The three failures are one item, an incomplete social preview image: the homepage has one, most inner pages do not. Of the 30 open items, 27 wait on the browser audit and 3 on a by-eye check of the main call to action.
The run on October 18 across nine sites: 142 checks passed, 5 warned, 3 failed and 30 are open. The failures are on the chart in red.
A standard you check against production is how a company keeps agent-built work from shipping half done.
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gianibrussich | ||||||||||||||||||||
| legalaimcp | ||||||||||||||||||||
| nyclaw | ||||||||||||||||||||
| nyrezone | ||||||||||||||||||||
| propertyappraiserpro | ||||||||||||||||||||
| rumblesimulator | ||||||||||||||||||||
| shankdit | ||||||||||||||||||||
| storiedandblessed | ||||||||||||||||||||
| westchesterwaterpro |
Swipe sideways to see all 20 items.
A personalized book has one chance to be right
The problem
Storied & Blessed sells a hardcover baptism book printed with the child’s name and the godparents’ names. On September 27 a buyer wrote at 11:07 asking about the godparents’ names. The book went to the printer at 12:01. Nobody was watching the inbox.
What I built
An order path that stops whenever it is unsure, and a watcher on the inbox that can stop the printer.
How it works
- Payment writes a record with a unique key before any slow work starts, so one order can make only one print. I added this after a paid order printed twice.
- Every edition is checked before it leaves. The boy and girl editions must match for each skin tone. Before I trust the check, I feed it bad pages and watch it fail.
- Books wait a set time before printing, longer than buyers are told, so a late change still lands in time. I made the wait longer after a change slipped through on October 10 and cost a reprint.
- Every 15 minutes a job reads the support inbox. A reply about an order that has not printed cancels the print job and holds the order.
- A person decides what happens to a held order. The software never guesses.
Before, a buyer’s message could sit unread until the book was printed. Now it is read within 15 minutes and the order is held.
When software acts on a customer's order, the safe default is to stop and ask, not to guess.
Unique record written first. One order, one print.
Edition match for every skin tone. A failed check stops here.
A buyer reply about this order?
Book printed and shipped.
Print job cancelled. A person decides.
The scarce thing is a person's attention. I treat mine as a budget.
Each of the four fixes does the same thing. It turns a question I used to answer by hand into a check that can fail on its own and tell me when it does. The jobs keep a count. The agents' instructions keep a signature. The sites keep a standard. The orders keep a hold box.
What is left for me is short, and it is written in a file my agents read before they start:
Anthropic's survey found that its staff expect far more from AI than the public does, because the two groups “have different experiences of what AI capabilities are.” This page is one of those experiences. Every number on it comes from a file I can open, and each fix made me needed a little less.
If your team is building ways to watch what its agents do, I would like to compare notes.
I can show the registry, the checks and the failures in a working session, with the real files open.