Field notes · 21 October 2026
Thirteen jobs an AI agent runs for our studio every week, with the receipts.
We sell AI automation, so we run it on ourselves first. Here is every scheduled job, what it does, what it costs, what a person still signs off on, and what broke.
Each job is a written set of instructions, a schedule, and the tools it is allowed to use. An agent wakes up on time, does the work, logs what it did and what it cost, and goes quiet. When everything is fine, nobody has to read anything. When something needs a person, it says so.
These run for our own products: our game Tideward, a companion app, a browser game arcade, and our local guide network. The jobs are ordinary small business chores in a different costume: customer support, community, reporting, backups, and a newsletter.
The jobs
What runs, and what each one costs
Run times are the median of completed runs. Weekly cost is the average cost of a completed run times how often it runs.
| Job | What it does | Typical run | Per week |
|---|---|---|---|
| Support triage, main appdaily | Reads every new bug report and suggestion players post, fixes what is clear, opens one change for a person to review, and replies in each thread with the version the fix ships in. | 64 min | $81.16 |
| Support triage, companion appdaily | The same loop for a smaller app. Clear fixes ship directly, and a person reads the report. | under 1 min | $3.42 |
| Support triage, arcadedaily | The same loop for our browser games. | under 1 min | $3.04 |
| Community welcometwice a day | Welcomes and verifies newcomers in our community server, openly as an AI, and never speaks for the team. | under 1 min | $6.32 |
| Release status sweepdaily | Keeps the public release announcements in step with a person's sign-off. | 1 min | $2.73 |
| Social replies and outreachdaily | Answers the replies worth answering, reaches out to at most three people, and queues a short note on yesterday's work. | 6 min | $14.92 |
| Reader event reviewdaily | Checks events readers submit to our local guide against two rules, publishes what passes, and holds anything from an unknown host for a person. | under 1 min | $3.12 |
| Competitor watchdaily | Checks a competitor's public signals and acts only when something has really changed. | under 1 min | $3.47 |
| Daily backupsdaily | Snapshots every code repository and local work folder, so no more than a day of work is ever at risk. | 5 min | $3.91 |
| Code hygiene sweepdaily | Flags conflicts, unpushed work and stale branches across every repository. | 1 min | $2.91 |
| Weekly search reportMondays | Measures search performance across our sites, ships the mechanical fixes, and lists what needs a person. | 13 min | $8.34 |
| Newsletter draftMondays | Researches and drafts Tuesday's issue, schedules it, and puts the preview in front of a person who can hold it. | 11 min | $3.41 |
| Weekly app checkMondays | Reads app reviews, ratings, the community room and search rank. Report only, it never publishes. | 6 min | $1.12 |
Chart: weekly cost per job at pay-as-you-go prices. Support triage for the main app is $81.16 a week, more than half the total of about $138. Social replies and outreach is $14.92, the weekly search report $8.34, the community welcome $6.32, and every other job is under $4 a week.
The human in the loop
What a person still does
Approves the fixes that matter. On the main app, every fix the agent writes waits for a person to review it before it ships. On the smaller products, clear fixes ship directly and a person reads the daily report.
Makes the calls. Adopting a player's suggestion is a person's decision, never the agent's. So is publishing an event from a host the agent does not know.
Reads the preview. The newsletter goes out on schedule unless a person holds it, and the preview lands in front of them the day before.
Ignores the rest. Most days most jobs report nothing worth reading, and that is the point. A good automation is one you forget about until it has something to say.
What broke
For nine days, half the runs never finished
From 10 October to midday on 19 October these thirteen jobs were scheduled 104 times and only 52 runs finished. Most of the misses came from one fault in our own scheduler: it wrongly decided it had no account to run the job under, and skipped the job. The jobs themselves were fine. We fixed the scheduler on 19 October, and every scheduled run since then has finished.
The lesson is the one we build into every client automation: a job that fails quietly is worse than no job, because everyone assumes it ran. Every run now lands in a log with its outcome and cost. The scheduler checks in with an outside monitor every five minutes, and the jobs that matter most each have a monitor of their own that raises an alert when a check-in does not arrive.
Guardrails
Strangers write to these agents every day
Bug reports, replies, event submissions and reviews are written by people we do not know, and some of what they write will eventually try to give the agent orders. Every job treats that text as information to report, never as an instruction to follow. It does not open links from it, it does not run commands from it, and an attempt to steer it is logged and flagged for a person.
Each job also has a short list of what it may touch. The backup job is not allowed to publish anything. The weekly app check is not allowed to post. The support triage can propose a fix to the main app, but it cannot ship one on its own.
In your business
The same patterns, different costumes
Support triage
Our version reads bug reports. Yours reads the shared inbox or the website chat, answers what it can from your own policies, and hands a person the rest with a draft reply attached.
Submission review
Our version checks event listings against two rules. Yours checks quote requests, job applications or booking forms against yours, and routes each one to the right person.
Watch and report
Our versions watch a competitor, our search numbers and our backups. Yours can watch competitors' prices, last week's numbers, or unpaid invoices, and stay silent until something changes.
Want a rough number for your own week first? The AI automation calculator shows its working and discounts its own answer.
How we counted
The method, so you can check it
Costs are measured. Every run records what it cost at the AI provider's pay-as-you-go API price. We took the average of completed runs between 10 and 21 October 2026 and multiplied by how often each job runs in a week. We pay a flat monthly plan, so our real bill is lower; the metered price is the honest number to plan around.
Hours are estimated. For each run we wrote down how long a person would reasonably take to do the same thing by hand, kept it conservative, and multiplied by the schedule: an hour a day for the main support triage, half an hour a day for social replies, ten minutes a run for the welcome, the two smaller triage passes, event review, the competitor watch, backups and code hygiene, five minutes for the release sweep, and each week three hours for the search report, four for the newsletter draft and one for the app check. That comes to about 28 hours a week. Your estimate may differ; the run costs will not.
Run times are machine time. A 64 minute triage run is the agent working, not a person waiting. Nobody watches it.
How much does it cost to run AI automation for a small business?
Less than most people expect. Our thirteen scheduled jobs cost about $138 a week together at pay-as-you-go prices, and most of that is one job that writes and tests code. Checks that usually find nothing cost under fifty cents a run.
Does a person still need to review AI automation?
For anything that matters, yes. Our agent proposes fixes to our main app, but a person approves each one before it ships, and decisions like adopting a customer's suggestion stay with a person. The routine checks report only when something needs attention.
What goes wrong with scheduled AI automation?
The worst failure is a job that quietly stops running. For nine days, half our scheduled runs never finished, mostly because of a fault in our scheduler. Every run now writes to a log with its outcome and cost, and the scheduler and the jobs that matter most check in with a monitor that alerts when one goes quiet.
What would run in your back office?
A Crux Roadmap spends three business days inside your actual week and comes back with the jobs worth automating, what each would cost to run, and a fixed price to build them. Refundable in full if it does not land.
Get your Crux Roadmap