HOME · WRITING · OPERATOR AGENTS
I run 10 AI agents for my businesses. Here's what they do.
The morning briefing, the inbox triage, the agent that audits the other agents — the actual fleet, the guardrails that make it safe, and what it honestly costs to run.
At 7am my phone shows an email titled with today's date and a one-line verdict on my businesses. It lists yesterday's sales and cash counts, the action items that aged overnight, and the two deadlines I would otherwise have remembered at 11pm. I didn't write it. I didn't prompt anything. An agent read the numbers while I slept and filed its report.
A few minutes later the growth agent has drafted the day's outreach and flagged a thread where someone finally replied. By 8am the triage agent has read four inboxes and left me a digest: three messages need me, one looks like fraud, the rest is filed. My inbox is triaged before coffee. That's the whole pitch, and it's not a pitch — it's my actual morning.
My position, before the details: I run this fleet across my own businesses — a juice bar franchise in Napa and the others — and I now build the same fleets for operators. So read what follows the way you'd read any builder describing their own work: useful, and biased. The evaluation checklist at the end is how I'd evaluate me.
What an AI agent actually is, in operator words
Strip the buzzword: an agent is software that does a recurring job for you on a schedule and reports back. The difference from ChatGPT is that you don't sit down and prompt it. It fires at 7am whether you remember it exists or not, does one defined job — read the numbers, triage the inbox, draft the replies — and lands its output where you already look. If you've ever paid a bookkeeper to send you a weekly summary, you already understand the shape. The new part is that the wage dropped to almost nothing and the summary shows up every day at the same minute.
The fleet, one job at a time
Ten agents, described exactly as they run today — genericized only enough to keep account details private:
- Morning briefing (daily). Reads yesterday's cash counts, sales, open action items, and upcoming deadlines. One email, priority-ordered. This is the agent I'd rebuild first if everything burned down.
- Inbox triage (twice daily, four inboxes). Separates needs-me from noise, flags fraud-shaped messages, drafts replies for what's worth answering. The drafts wait in the outbox; nothing sends itself.
- Growth drafts (daily). Drafts outreach and content posts from a queue I control, and watches for replies to earlier outreach so warm threads don't go cold while I'm on the floor.
- Fleet reviewer (weekly). Reads the other agents' output for the week and proposes tuning — a prompt to tighten, a check to add, noise to cut. I approve or reject each change. The system that watches the system still reports to a human.
- Ops pattern-miner (weekly). Digs through the store's operational data — counts, deliveries, waste — for patterns an owner stops seeing when every day looks like the last one.
- Product signal-miner (weekly). Same idea, pointed at product and customer signals: what's gaining, what's fading, ahead of the monthly numbers.
- Doc auditor (scheduled). Compares the operating docs against reality and lists which documents have quietly become fiction. Every operation has these; mine just get caught.
- Visibility report (monthly). Checks where my businesses actually appear in AI answers and search results against a fixed set of buyer questions, scored against last month.
- Search tracker (weekly). Watches search queries, positions, and pages across the business sites and reports the deltas worth acting on.
- Local-SEO watcher (weekly). Tracks local-search rankings and listings for a local service business I run, and names the specific fixes.
Notice what's not on the list: nothing customer-facing, nothing that touches scheduling or payroll, nothing that moves money. Agents earn wider scope slowly, and some scope they should never get.
The guardrails — the part that actually matters
Every operator I talk to has the same fear, and it's the right fear: AI doing something on its own, to a customer, in your name. So here is the design rule the whole fleet runs on: agents draft, they don't send. They propose, they don't apply.
- The triage agent writes the reply to the catering inquiry — I hit send.
- The growth agent drafts the outreach — it sits in my outbox until I approve it.
- The reviewer agent proposes changes to the other agents — nothing self-modifies; I approve each one.
With that boundary in place, the worst case of a bad agent collapses from "runaway AI emailed my customers" to "I deleted a mediocre draft." That's a failure mode you can live with while the fleet earns trust. And the weekly review isn't ceremony — models change, my businesses change, and an unaudited agent drifts out of date like any unread report. Human-in-the-loop isn't a compliance checkbox in this design; it's the reason the design works.
What it honestly costs and takes to run
Three honest lines, because this is where agent content usually goes soft:
- The running cost is model usage plus tooling subscriptions, billed to your own accounts. For a draft-and-report fleet like mine it's a rounding error next to one labor shift. I won't quote you an exact monthly figure because it depends on fleet size and how much each agent reads — but it's the smallest line in the math.
- The real cost is tending. Prompts drift out of date. A report format changes upstream and an agent starts reading garbage. A job that mattered in March stops mattering in July. Somebody has to notice, tune, retire, and add — that's a recurring hour, not a one-time build.
- A technical owner can absolutely DIY a starter version. If you're comfortable with scheduled scripts and API keys, start with a morning briefing and see what it changes. The build-for-you path exists for operators who want the fleet without becoming the maintainer.
What a starter fleet looks like
Single location: four agents, in this order — a daily owner briefing pulled from the reports you already have; inbox triage with drafts; a review-and-inquiry responder that drafts every reply in your voice; and a vendor-email watcher that catches price-change notices and substitution alerts before they cost you money. That set absorbs the administrative hour most owners run before open.
Multi-unit: the same four, plus a consolidation layer — one briefing across stores with per-store exceptions ("store two's cash variance is the outlier today"), and triage that understands which messages are store-level versus brand-level. The fleet's value compounds with unit count because the 6am hour scales with stores and your morning doesn't.
How to evaluate any agent builder — including me
"AI agent builder" is 2026's least-regulated job title. Whoever you talk to — an agency, a freelancer, me — run the same five checks:
The honest close
I built this fleet for myself because I was the bottleneck in my own businesses, and the 6am hour was the first thing worth automating. I now build the same fleets for operators — scoped inside the audit, built as a fixed-quote project, tended on the retainer. The checklist above is how I'd evaluate me or anyone else. Run it hard, and if a builder fails it, keep your money — including from me.
Wondering which agents would pay rent in your operation?
That's exactly what the $2,500 AI Operations Audit answers — every AI opportunity in your operation priced by payback, agents included. Fleets are sold inside the existing engagements: audit → build → retainer. No new tier, no agent-shaped upsell.
See the Operator Agents pageDRAFT-ONLY GUARDRAILS BY DEFAULT · I RESPOND WITHIN 24 HOURS