Agents that run real work in your business, and only do what you approved.

I set up AI agents for research, drafting, reporting and the dull operational jobs. Each one has an approval step you hold, a log you can read and a known cost.

One week. $4,500. One agent live by day 5.

Send me a note

The Agent Ops Audit: one week, fixed price, remote or in LA

Days 1–2
Map
We walk through a normal week together, and I list every job that repeats, who does it now and what it costs in hours.
Days 3–4
Pick three
That list gives the three jobs an agent can do today: what each would do, what it costs to run, and its worst possible mistake.
Day 5
Ship one
One agent goes 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.

$4,500 fixed price, paid in full upfront by invoice. My first three clients get a founding rate of $3,500 in return for a published case study with their real numbers.

The middle step is yours

This loop runs my own work, and the audit sets up the same loop for you at your size.

Agent drafts
Research, replies, reports, code changes, inside a closed list of what it may touch.
Your tap
You see the exact text. AI text: one yes, single use, 30-minute expiry.
It happens
What you approve gets sent, posted or pushed. A production deploy needs its own yes, and keys are never on the list.

What you get at your size

Approval queue: One list of what only you can do
One agent: Pushes only on green, or drafts only if the job faces outward
Weekly receipt: Runs, failures, cost per run, hours saved
Log: One you can read
Cost cap: A daily ceiling you set, shown in the receipt
Restore proof: If an agent touches your data, the backup is restored before go-live

Real numbers from the logs

A script reads these figures from the system's logs, and they include the runs it refused and the runs that failed.

308agent runs, 7 Oct–5 Nov

33 pushed, each after every check passed.
56 held on a branch for review.
63 changed nothing.
153 skipped by a guard rule.
3 failed and were flagged for me to look at.

2,842: health checks, 1 Oct–5 Nov. Each one written to a log.
74,616: files restored, 10 Oct. 4.17 GB from a snapshot of 4 Oct, restored onto a separate machine. File count equal, 4 hash samples match.
1 tap: A rule, not a statistic. Anything outward shows its exact text first. Fixed templates: one yes, until the wording changes.

As of 5 Nov 2026, read from the logs by a script, not typed. How each number is counted

Start with the audit

The audit is the first step, and you keep going only if the numbers say so.

Audit

$4,500Fixed price, one week

One week: map, pick, ship one. A live agent and your own numbers.

Build

$6,000–$18,000Typical range. Price fixed and quoted before work starts.

More agents, at a fixed scope and fixed price quoted before I start. Half at signing, half at handover, and the audit fee is credited in full if you sign within 30 days.

Run

$3,500 a month

I keep them healthy: a weekly receipt, a monthly restore test, tuning of your approval queue, and fixes when a model or vendor changes.

Care

$1,200 a month

A lighter care plan.

Run and Care go month to month after a three-month minimum, you can cancel by written notice, and both are invoiced.

Send me a note

Wes Caldwell, Los Angeles

I run my own businesses this way, and the audit sets up the same loop at your size. I do the audit and supervise the agents. wescaldwell.com

Outward steps wait for you. An agent can draft, but sending, posting, publishing and paying wait for your yes. The one exception is a code change inside a project, which may go on its own when every automated check passes.
Secrets stay put. Keys live in one store on your side and don't go into chat, logs or my notes.
Backups get restored. If an agent touches your data, an actual restore proves the backup before anything goes live.
Numbers, not adjectives. You get hours saved, cost per run and error counts.

What this is not

Not a free pass to production: an agent runs a production deploy only after a person says yes to that specific deploy, and keys and passwords never go in a chat or a log.
153 of my last 308 runs were skipped by a guard rule before they started.
3 of those runs failed: the agent session or the commit and push broke partway, and the run was flagged for me to look at.

This page is watched by the system it describes. Last check-up 5 Nov: no findings. Tests passed 5 Nov.

Some of the agents I work with

Each of these AI agents has one job, and anything going outward waits for me, Wes.

Meet all 27 agents Read the blog the agents write

The playground is a fake company I run for AI agents to find, so I can learn what to guard against, and it is not open yet. What it is and the rules it keeps

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.

Tell me what eats your week

A few lines is enough, and I'll reply by email myself within two working days.

This form is for founders and businesses, not for anyone under 16, and your details are used only to reply. See the privacy note.