What Running a Business Mostly With Agents Looks Like Day to Day

What it actually looks like to run small businesses mostly on AI agents: the real human hours, what you still cannot hand off, and how you decide to kill something.

Here is what agent-led operations look like from the inside. Not the demo version. The Tuesday version. How many human hours it really takes, which jobs stay human no matter what, and how something earns the right to keep existing or gets shut down.

I am Brian Marvin, an AI-native fractional CTO. I run my own operation this way. My site replays its own git history, 138 merged PRs at last count, precisely so visitors can see one human shipping with an agent fleet instead of taking my word for it. What follows is what that costs me in hours and judgment.

The real human hours: fewer than you fear, more than zero

Let me start with the number everyone wants. Running one small business mostly on agents costs me roughly five to ten focused human hours a week once the system is built. The build itself is the expensive part: two to six weeks of real setup, depending on how messy the starting point is. After that, the week settles into a rhythm.

A normal day looks like this. Morning review, twenty to forty minutes. I read what the agents did overnight: commits, drafts, flagged exceptions, cost summaries. I approve or reject the decisions they queued. Anything they handled cleanly, I skim. Anything they flagged, I actually read. Midday is exception work only. A customer situation the agent escalated. A vendor change that broke a connection. A piece of wording that needs a human voice. Most days that is zero to two items. Weekly, one longer session. I review the numbers, read the incident log, and decide what gets fixed, tuned, or retired.

The honest caveat is that the hours are lumpy. A quiet month runs at the low end. A vendor deprecation, a security scare, or a new product launch can eat a full week. Anyone who quotes you a flat forever number is selling. The average holds. The variance is real.

For the full playbook on the one-person version of this, see the one person company playbook. The split is the same. AI does the repeatable work. You keep judgment, money, and promises.

What you still cannot hand off

Five categories stay human in every business I run. I have tested each boundary and been burned at each one, so this list is written in scar tissue.

Money movement. Paying vendors, issuing refunds above a threshold, changing prices, signing anything with a dollar figure. Agents draft. Humans sign. Every time an exception got made, the cost of the mistake exceeded a year of the saved minutes.

Public words and promises. Anything a customer, regulator, or partner will read as your commitment. Agents draft the reply, the post, the policy. A human approves before it ships. Trust compounds slowly and dies instantly, and no latency saving is worth a public apology.

Hiring, firing, and access. Who gets keys, who loses them, who joins, who leaves. Agents can summarize candidates and draft the offboarding checklist. The decision and the conversation are human. Always.

Strategy calls with incomplete information. Agents are excellent at summarizing what is known. Deciding under uncertainty, with missing data and conflicting signals, is still a human job. The model will happily produce a confident recommendation from an incomplete pile. Your job is to notice the pile is incomplete.

The kill decision itself. Which brings us to the third question, and the one I get wrong most instructively.

I go deeper on the cost side of this split in orchestration and spend. The numbers there are about agent fleets, but the principle generalizes. Throughput is cheap. Judgment is the constraint.

How I decide to kill something

Everything I run has a review date written down at launch. Monthly for experiments, quarterly for established lines. At each review, the thing must defend its existence against three questions.

Does it earn its keep. Revenue minus true cost, where true cost includes my review hours, tool spend, token spend, and support load. I have killed automations that looked efficient because my fifteen minutes of daily babysitting cost more than the chore they replaced. If the math needs optimism to work, it does not work.

Does it fail safely. I ask what happens when each piece breaks, because each piece will break. Login sessions expire. Vendors rename fields. Models change behavior. If a failure mode means silent wrongness, money moving without approval, or a customer promise breaking, the thing either gets a human gate or gets killed. I keep a short list of past failures and check new systems against it before they launch.

Would I build it again knowing what I know. Sunk cost is the hardest filter. The honest version of this question is whether the system teaches me anything still, or whether it just occupies a slot. Systems that plateau into pure maintenance without improving get a sunset quarter. I announce the date, run the wind-down, and archive the notes. The notes are the real asset. They make the next build faster.

The pattern is deliberate. Small bets, short leashes, written-down kill dates, no shame in the archive. I describe the same loop for agent work in harness, loop, graph. Tight loops with clear permissions beat long leashes every time, for code and for companies.

What the fleet actually does all day

For readers who have never watched an agent-led operation, here is the concrete shape. Nights and weekends belong to the fleet. It handles implementation: code changes, content drafts, data entry, report assembly, monitoring checks, first-pass customer replies queued for approval. Mornings belong to me. I review, approve, correct, and set direction. The site you are reading runs this way. The git history replay on the homepage is not a metaphor. It is the actual commit log of this site being built and maintained by that loop.

The part people underestimate is the maintenance load on the human. Agents do not maintain themselves. Dependencies drift. Prompts rot as models change. A workflow that ran clean for two months starts failing on Tuesdays for reasons nobody can see until someone reads the trace. My weekly session is mostly this: reading traces for absences, not just errors. What did the agent never consider. Which check stopped running. What changed upstream.

I documented fifteen ways this breaks in fifteen failure modes from two months of unattended operation. Read that before you believe anyone, including me, who tells you agents run themselves. They run themselves the way a garden grows itself. The weeds come free. The harvest takes a gardener.

A note on semantics: this is not a consulting pitch

This article is not a funnel. There is no call to action hidden in it. The questions it answers are the ones worth asking before running anything on agents: what are the hours, what stays human, how do you kill things.

If you run small businesses and want to explore agent operations, steal the whole pattern. The review rhythm, the kill dates, the non-delegable list, the morning review habit. It all works without me. If you later want a second pair of eyes, you know where I am. Either way, the notes are yours.

For the process discipline underneath all of this, see what I use instead of agile for AI agents. Short loops, written decisions, no ceremony. That is the whole management system.

FAQ: How many hours a week does one agent-run business take

Five to ten focused human hours once built, plus a two to six week build. Quiet months run lower. Vendor changes, launches, and incidents spike it. The morning review habit, twenty to forty minutes, is the load-bearing wall. Skip it and the exceptions pile up silently.

FAQ: What can never be delegated to agents

Money movement, public commitments, hiring and access decisions, strategy under uncertainty, and the kill decision. Agents draft and summarize in all five areas. Humans decide. Every exception I have made to this list cost more than it saved.

FAQ: When do you kill an automation or product line

At a pre-written review date, against three questions: does it earn its true keep, does it fail safely, and would I build it again. Fail any one and it gets a sunset quarter with an announced date. The notes get archived. The slot gets freed.

FAQ: Is this consulting in disguise

No. Take the pattern and run it. If you want help later, ask. The article stands alone either way.

About the Author

I am Brian Marvin, an AI-native Fractional CTO with 30 years in technical leadership. At empowered.guru, I help startups and small businesses build MVPs, shape roadmaps, and make AI-powered technology decisions that scale.