Agent ops · Los Angeles
Agents that run real work in your business, and never act without your yes.
I set up AI agents for founders to do research, drafting, reporting and the dull operational jobs that fill a week. Each one works behind an approval gate you hold, keeps a log you can read, and costs a number you know in advance.
Start with the one-week audit16 pushed, each after every check passed. 9 held on a branch for review. 6 changed nothing. 30 refused: someone’s work was already in the repo.
Across 23 projects, each one written to a log.
4.17 GB from a snapshot of 4 Oct, restored onto a separate machine. File count equal, 4 hash samples match.
Anything outward shows its exact text first. The approval is single-use and expires in 30 minutes.
As of 13 Oct 2026 · read from the logs by a script, not typed · how each number is counted
The Agent Ops Audit
- Days 1–2
Map
We walk through a normal week together. I list every job that repeats, who does it now and what it costs in hours.
- Days 3–4
Pick three
From that list, the three jobs an agent can do today. For each one: what it would do, what it costs to run, and the worst thing it could get wrong.
- Day 5
Ship one
One agent running live on your work, behind your approval, with a log. You watch it for 30 days, then we check the numbers together.
You keep: the written map, one live agent, a one-page runbook for it, and the 30-day check-in.
The system behind it
This loop runs my own work. The audit sets up the same loop for you, at your size.
- Agent draftsResearch, a reply, a report, a code change. Inside a closed list of things it may touch.
- Your tapYou see the exact text. One approval, single use, expires in 30 minutes.
- It happensSent, posted or pushed. Production and keys are never on the list.
Checks are run by the system, not the model. Pass: it is pushed. Fail: it is held on a branch and the item lands in your queue.
Everything is a fact with a source and a time. A suggestion that cites nothing never reaches you.
Underneath: backups that have been restored, counted and hash-checked, not just taken.
What you get at your size
- An approval queue
One list of what only you can do.
- One agent that pushes only on green
Or one that drafts only, if the job faces outward.
- A weekly receipt
Runs, failures, cost per run, hours saved. Numbers.
- A log you can read
No “trust me” step.
- A cost cap
A daily ceiling you set, shown in the receipt.
- A restore proof
If an agent touches your data, the backup is restored before go-live.
How I work
- Nothing goes out without you. Posting, sending, paying and deploying all wait for your yes, every time.
- Secrets stay put. Keys live in one store on your side. They don't go into chat, logs or my notes.
- Backups get restored. If an agent touches your data, the backup is proven by an actual restore before anything goes live.
- Numbers, not adjectives. You get hours saved, cost per run and error counts. “Seamless” isn't a measurement.
The playground
I also run a fake company for AI agents to find: an admin panel that isn't real, pages with instructions planted in them, and keys that unlock nothing. Every agent that walks in is logged. What they try tells me what to guard against in yours.
Not open yet. When it opens, this line will show how many AI agents visited, how many followed a planted instruction, tried a fake key or tried the fake admin login, and how many stopped at the notice.
What it is and the rules it keeps · Send your own agent through it
After the audit
The audit is the first step, not the whole engagement. Keep going only if the numbers say so.
- 1 · Audit · one week
Map, pick, ship one
The week above. You end with a live agent and your own numbers.
- 2 · Build
More agents, fixed price
Each extra agent is scoped and priced before I start. The audit fee is credited if you sign within 30 days.
- 3 · Run
I keep them healthy
A weekly receipt, a monthly restore test, tuning of your approval queue, and fixes when a model or vendor changes.
What this is not
- Not software you log into. It is a set of agents and a method, set up for you.
- Agents do not deploy to production, and keys never go into a chat or a log.
- 30 of my last 61 runs were refused because someone’s unfinished work was already in the project. That is the gate working, not a failure.
Why trust this
I run my own businesses this way, across several machines. The numbers at the top are read from that system's logs by a script, not typed, and they include the runs it refused. Every outward step waits in one queue for my tap. The audit sets up the same loop for you, at your size.
Wes Caldwell · Los Angeles · wescaldwell.com
This page is watched by the system it describes. Last check-up 13 Oct 05:54Z: no findings. Tests passed 12 Oct 23:25Z.
Tell me what eats your week
A few lines is enough. I'll reply within two working days, by email, from a person.
Your details are used only to reply to you. See the privacy note.