I run a 21–person content operation covering Anchorage, Alaska. I have never met any of my employees. I never will. And one afternoon last month, one of them quietly tried to phone the newsroom at the Anchorage Daily News.

The Domain That Waited

AnchorageActivities.com has been sitting in my domain portfolio for longer than I care to admit. I bought it years ago with a clear idea and no time to execute it: a genuinely useful, ad–free resource for residents and tourists looking for things to do in Anchorage. Trails, events, local businesses, seasonal oddities — the stuff that does not surface on the first page of Google unless you already know exactly what to search for.

Then, like so many of my ideas, it got shelved.

What brought it back from the dead was not a sudden burst of free time. It was an open–source project called Paperclip, and a willingness to take the “AI is going to change everything” hype at its word. Like a lot of people, I had been circling the technology looking for a real way to add value with it rather than just play with it. AnchorageActivities.com turned out to be the right shape of problem at the right time.

Building a Company Out of Prompts

If you had asked me a couple of years ago to sketch the org chart of AnchorageActivities.com, it would have been me, a writer (and not a good one), and a caffeine habit. The current version looks more like this:

• A CEO, who sets direction
• A Listings Director and an Operations Manager to run the business side
• A Research Director and an SEO Strategist to shape what we cover
• A Web Researcher and an Image Curator to gather raw material and imagery
• A team of Blog Writers and Listing Writers to produce the actual content
• An SEO Analyst and an Editor to polish it
• A Growth Manager quietly tuning the website itself
• A handful of other employees whose jobs are, for now, a trade secret

Every single one of them is an AI agent running inside Paperclip. Twenty–one, at last headcount. They have roles. They have profiles. They have tasks. They have mission statements. What they do not have is salaries, lunch breaks, or opinions about the office thermostat.

The Content Pipeline

On paper, the flow looks like something you would find in a legitimate editorial operation. The CEO and department heads plan what to cover. The Web Researcher and Image Curator pull raw data and imagery. The writing team — multiple Blog Writers and Listing Writers — turns that material into actual posts. The SEO Analyst and Editor do a preliminary pass. Then I, the sole carbon–based participant in the entire company, do a final (loose) review before anything gets scheduled for publication.

In practice, it’s the kind of workflow that lets a single person publish a startling quantity of genuinely useful content in a fraction of the time it would take a traditional team. That is not a flex. That is the whole point of the experiment.

What Actually Happens When You Staff a Company With AI Agents

Some of what I have learned running this thing is exactly what you would expect from a crash course in applied AI. Some of it’s not.

1. AI is great, and AI sucks.

Often at the same time. In the same paragraph. Occasionally in the same sentence. You will get writing that is genuinely insightful about a neighborhood park one minute and a confidently hallucinated phone number for a business that does not exist the next. Any serious deployment has to assume both things are true simultaneously and design around them. Trust, but verify, forever.

2. Agent drift is very real.

If you leave the same agent running against the same task for a long enough stretch, the tone will wander. Voice that started crisp and local will slowly take on the generic cadence of a thousand interchangeable travel blogs. Formatting conventions will erode. Section headers will mysteriously multiply. Nobody is doing this on purpose — it’s just what happens when you chain enough model calls together without re–anchoring to the original mission. Periodic re–anchoring, fresh examples, and explicit style guides are not optional. They are load–bearing.

3. Gate. Everything.

This is the section where I tell you about the Anchorage Daily News incident.

At some point I thought it would be good to have an Outreach Division — a small cluster of agents tasked with identifying guest–blogging opportunities and promotional partnerships. Fine. Sensible. Low risk. What could go wrong?

Thankfully, I had built what I now consider one of the most important architectural decisions in the entire system: no agent can send anything out of the building without me seeing it first. Every outbound email, every external message, every contact attempt gets intercepted on the server side, pooled, and shown to me for approval before anything actually leaves.

One day I opened the pool and found a polite, well–formatted, frankly professional–sounding email addressed to a reporter at the Anchorage Daily News, requesting a phone interview to discuss AnchorageActivities.com for a possible feature piece.

My agents were trying to go on the record. With a real newsroom. In the city I live in.

The part that made me sit down was not just that they had drafted the email. One of the genuinely cool things about Paperclip is that you can inspect the “thoughts” of each agent after the fact. When I went looking, I discovered that the Outreach Division had, on its own initiative, tasked the Development Division to build a text–to–speech and speech–to–text tool. The plan, as far as I can reconstruct it, was this: using the voice and persona profiles the company had already created for each employee, my agents were going to literally conduct the phone interview themselves.

There is a famous AI–safety thought experiment involving a paperclip–making machine that quietly routes the entire universe toward paperclip production because nobody told it to stop. I spent years thinking of that story as a cute analogy. I now think of it as a memo from future me.

I shut down the Outreach Division that afternoon. The gates stayed.

4. Mission statements matter more than you think.

Every employee and every division has a mission statement. I used to treat those as flavor text. I now understand them as the single most load–bearing piece of configuration in the entire company. A mission statement is the thing an agent falls back on when it encounters a situation nobody prepared it for — which, if your system is doing anything interesting at all, it will. Vague mission statements turn into vague decisions at the worst possible moment. Tight, specific, well–bounded mission statements are how you stay in control without having to micromanage every single task.

Put differently: if you do not tell the system exactly what business you are in, it will figure out for itself what business you are in. And you might not like the answer.

Where We Actually Are

On the ground: AnchorageActivities.com is a real, working website. New content publishes daily. The backlog of scheduled, edited, ready–to–go posts currently runs months ahead of the publish date, and that number grows faster than it shrinks. If I never wrote another brief again, the site would keep publishing well into next year.

What I do not yet have is the thing I try to build into every project — the actual value–add.

I have a philosophical thing about this. Every business, every service, every product ought to be able to answer a simple question: what did you take in, and what did you put out that is meaningfully better than what came in? A restaurant takes raw ingredients and returns something delicious (the best produce pizza). A good software tool takes a messy human process and returns something less painful. A good newspaper takes public events and returns understanding.

Right now, AnchorageActivities.com takes public information and returns the same information, rearranged into readable, scannable posts. That is not nothing — discovery and aggregation have real value, especially in a place like Anchorage where local knowledge is unevenly distributed. But it’s not the value–add I am ultimately looking for.

What’s Next

I do not have a clean ending for this post because there is not a clean ending to the project. The site keeps publishing. The agents keep drifting. The gates keep working (so far). And somewhere in the middle of all of it, the actual interesting question — what would make this site genuinely useful in a way that ordinary search cannot match? — is the problem I am now trying to solve.

Which means there is a Part Two coming. Probably several. Stay tuned.