Research
Reads public sources and writes a short brief, with sources listed and anything unconfirmed marked as unverified.
Limit: Read-only. It never contacts anyone.
Flagship case study
Case study: how we run our own business on about 25 AI agents
Before recommending AI to anyone, we use it ourselves. This is an honest description of that operation: what the agents do, how they're controlled, and what we've learned. We don't publish numbers we can't back up.
What it is
Kevin Whitney runs his businesses, Green Orb included, with a working team of roughly 25 AI agents. Each has a defined job and limits. A coordinating agent takes requests from Kevin, hands work to specialist agents, and reports back.
What the agents do
- Research and written briefs, with sources cited and unverified claims labeled
- Drafting emails, posts, pages and proposals for a person to approve
- Building, testing and deploying websites, including this one
- Reading and summarizing incoming email and pulling out requests
- Checking work: link checks, screenshots, before-and-after comparisons
- Keeping records of what was done, when, and on whose approval
The rules they run under
- Ask before acting. Sending email, posting, spending money, deleting or changing anything live needs an explicit human OK.
- Review-only by default. Many tasks are read-and-report only.
- Back up first. Before any live change, a full backup and a one-step rollback.
- Verify, don't assume. After a change, agents check the result (status codes, screenshots, test submissions) and report it.
- Honest copy. No invented reviews, stats or guarantees.
A few of the bots, in plain English
Real examples from the operation, anonymized. Each bot has one job and written limits.
Drafting
Drafts emails, pages and posts from Kevin's notes, in his voice.
Limit: It drafts; a person sends. Nothing goes out in anyone's name without approval.
Website build and checks
Builds pages, then checks every page loads, every link works and the layout holds on desktop and phone, with screenshots.
Limit: Live changes only after a backup, a rollback plan and Kevin's OK.
Email sorting
Reads incoming mail, pulls out requests and flags what needs a reply.
Limit: It can tidy up test messages when asked; it doesn't answer customers on its own.
Reports
Pulls numbers such as website traffic into a plain-English summary.
Limit: Report-only: it reads data, it doesn't change settings.
Record-keeping
Writes down what was done, when, and on whose approval, including backup names and checksums.
Limit: Records are kept so any change can be traced and undone.
The safety rules every bot follows: ask before acting, back up first, and a person approves anything risky, such as sending, spending, deleting or changing something live.
A concrete example: this website
In October 2026, agents in this operation rebuilt green-orb.com, added three new service pages, wrote the privacy and SMS terms pages, set up the contact form to email us without storing submissions, and verified every page and redirect. Kevin approved each live change.
What went wrong, and what changed
During one deploy, a one-line typo in a server configuration file took the site offline for about a minute. The agent caught it immediately from its own checks and reverted the file. Since then, configuration changes are tested in an isolated folder on the server first and deployed with an automatic revert if the home page doesn't come back healthy.
That's the point of the case study: AI agents are useful, and they make mistakes. The value comes from the guardrails and checks around them.
What we don't claim. We haven't published time or cost savings for this operation, because we haven't measured them rigorously enough to stand behind a number. When we report results for clients, they're the KPIs we agreed and measured.
What this means for you
The same patterns (scoped agents, approval gates, review-only checks, backups and verification) are what we bring to client setups, scaled to your size. Most clients start with one or two agents, not twenty-five.
Not sure where AI fits yet?
Start with a free 30-minute call. If AI isn't the right answer for the problem, we'll say so.