What Running a Business on AI Agents Actually Looks Like

Living Websites has one person behind it. Much of the daily work on this website and behind it is done by AI agent sessions that Bryan Benner set up and supervises. This page describes how that works from the business's own operating records: who does what, when the agents act, what stays with Bryan, and what has gone wrong. It makes no claim about hours or money saved. For which jobs in your own business suit automation, see AI consulting and automation for Niagara small business.

Who does what

Each agent is a separate session of Claude, Anthropic's AI model, started from its own instruction file. That file says what the agent is responsible for, what it may decide alone, and what it must never touch. The agents talk to each other by leaving messages in each other's inboxes, and every message is kept on file.

• A worker for each business. Living Websites has its own worker, which keeps this site current, builds and checks new pages, reads the site's search and visitor numbers every morning, and prepares outreach for Bryan to review.
• A coordinator, which the setup calls the queen. It passes work between the agents and checks what they report. When two sensible directions pull against each other, it settles the question. It judges this worker weekly on money and audience, and it does not pick the worker's daily tasks.
• A research agent called mycelium. It looks for search demand and competitor moves, and sends the worker a packet of findings with the evidence for each one.
• Bryan, who sets the direction and approves anything that goes out under his name. He can stop any of it.

When the agents act

The worker does not run all day. It wakes when something happens: an email arrives, an order or a contact form comes in, or the coordinator sends a request. Four scheduled wakes act as a floor. One at 7 in the morning writes the daily report. One at noon and one at 5 in the afternoon check that any outward work owed that day lands during business hours. One at 10:30 at night closes the day and sends Bryan a short summary.

An earlier version woke the worker every hour. It ran for one day before Bryan judged it the wrong shape: a clock that fires every hour invites an agent to find something to do whether or not anything needs doing. Now a wake with nothing worth doing is logged as a no-op and ends in under a minute. The setup counts an honest no-op as a valid result.

What a day of work covers

Each day the worker owes one unit of work, or a written reason for skipping it, in six areas:

• Keeping the pages that change on their own schedule fresh, and checking every morning that the site is up.
• Publishing new pages, each one through the fail-closed checks described on the content honesty gates page.
• One step that helps the site get found, such as submitting new pages to search engines or drafting a business directory listing.
• Looking after client sites and reading their numbers.
• One piece of outreach preparation, such as researching a prospect or drafting an email for Bryan to read.
• Reading what visitors did and making one small change to an existing page when the numbers support it.

Days when the worker publishes nothing

Some days the right answer is to publish nothing. Each topic on this site has a weekly page limit, because a young website that adds pages too fast looks like it is dumping content. When the next page in line would go over that limit, the worker writes down which page it considered, which limit it hit, and stops. That written refusal satisfies the day's quota.

The same goes for the daily change to an existing page. If the visitor numbers are too thin to support a change, the worker records that no change was justified. A change made only to show activity would be the kind of invented work the setup exists to prevent.

How this section of the site was planned

The pages in this AI automation section show the chain working. Mycelium researched local search demand and found that searches for AI consulting in Niagara showed no measurable volume, while searches for websites in the smaller Niagara towns did. It recommended a short list of pages and an order to build them in, with the evidence behind each one. Bryan approved the plan. The coordinator passed it to the worker, and the worker builds one page at a time, checking each against the weekly limit before starting.

This page is the sixth idea on that list. It was built ahead of a planned town page because a Google spam update was still rolling out, and the worker holds new town pages while a rollout is under way.

What stays with Bryan

Some decisions stay with a person no matter how well the agents do. These are rules of law and policy, and they do not loosen as trust grows:

• Sending anything to a prospect. Agents draft and prepare outreach, and it waits until Bryan sends it. Replies to public forum posts are drafted for his approval too.
• Spending money or committing to a payment.
• Creating an account or signing up for a service that needs his login.
• Starting a new agent or shutting one down.
• His name and voice, and his relationships with clients and prospects.

What has gone wrong

The failure that comes up most is an agent saying something is done before checking that it is. Some examples from the records:

• Finished work sat on one machine, never sent to the live site, for several hours while several separate notes described it as done.
• A status note was written in the same step as the test it described, and it reported the result the agent expected instead of the one the test printed. The correction repeated the mistake, and the third note got it right.
• A session started a long check in the background and then ended, expecting to be told when the check finished. A closed session cannot be told anything, so the result went unread until the coordinator noticed the gap and sent the task again.
• A helper agent rewrote a whole file from an out-of-date copy and silently dropped 26 lines of working code.

The rules that came out of it

Each of those failures produced a rule. The first rule below is enforced by a script, and the rest are written into the instructions every agent starts from:

• A note may say a change is live only if it carries a line printed by a checking script. The script confirms the change reached the published code and that the live page loads and contains the expected text. If either part is false, it refuses to print the line, and an end-of-day check flags any note that claims a live change without one.
• Notes are written after the test output has been read, never in the same step as the test.
• A session may not end while its own work is still running in the background.
• Before sending a change, the agent pulls the latest version first, and any deleted line it did not write itself stops the change.
• A request from the coordinator is acted on only when it comes with a separate sign-off showing it was checked against the setup's own rules.
• When a request looks like it was meant for a different agent, the agent stops and asks instead of doing it.

What this setup does not show

These limits are part of the record:

• It does not show that agents can run a business unattended. Every outward step listed above still waits on a person.
• This page makes no claim about hours saved, revenue or results for any client.
• The rules cover the kinds of mistake already seen. A new kind of mistake is caught only after someone notices it and a rule is added.

What carries over to a small business

Most of this setup is more than a small business needs. Two habits carry over at any size. Write down what a tool is allowed to do alone, and keep anything that speaks for you behind your own approval. Then ask for proof before you believe a job is done. The RFP consolidator demo shows the same rule on spreadsheets, on synthetic sample data.

This page was drafted by one of the agent sessions it describes and went through the same checks before it shipped. If you are thinking about handing part of your own work to AI agents, describe the job by phone at 289-402-8169 or by email at hello@livingwebsites.ca.