I asked Codex to send a fresh agent to my newsletter and help a stranger decide whether to subscribe.
The reader I gave it was someone who uses AI every day but doesn’t code. I wanted the agent to explain who AAC is for, pick three articles to start with, and work out what a subscriber would get.
It got most of that right. It understood the audience and chose a sensible reading order. Then it reached the questions about courses, support, and paid access, and couldn’t verify enough to give a complete answer.
That was the useful part of the experiment. The agent could explain my work, but some of the details a reader would need before paying were still out of reach.
In This Edition
What an agent learned about AAC from the public site.
The three articles it chose for someone who doesn’t code.
Why finding an announcement didn’t tell it what was available.
A prompt to try on your newsletter, consulting site, or public project.
Subscribe to All Agents Considered for practical experiments with AI agents, including the ones that send me back to inspect my own work.
Give It a Stranger’s View
I’ve already asked Hermes what my files say I do all day. It described me as an AI Workflow Field Reporter With a Side Cart of Websites , which is still a ridiculous thing to put on LinkedIn.
Hermes had an advantage in that exercise. It could see the work behind the articles, including notes and project files a reader would never encounter.
This time I wanted the public version. Someone looking at my newsletter doesn’t know which plans I’ve abandoned or what I meant when I wrote the About page. An agent helping that person should have to work from the same evidence.
A few days ago, I had Codex run the task through a subagent, a separate agent conversation that received none of our earlier discussion. It got the public URL and instructions to use public sources, avoid local files, and say when it couldn’t verify something.
The request came down to this:
I use AI every day, but I don’t code. Should I subscribe to All Agents Considered, and which three articles should I read first?
I also asked what free and paid subscribers get, whether it could confirm any courses or support, and who should skip the publication. Each answer needed sources.
The agent reported that it used no local files. Its tools still had workspace access, so this depended on following the instruction. The test gave it a fresh conversation, not a physically separate computer.
We saved its answer before sending any feedback. That gave me a record of what it found without being coached toward my description of AAC.
It Found the Right Readers
The agent described AAC as practical writing about getting more control over everyday AI work. It picked up the recurring subjects: agents, local models, memory, files, automation, and the checks that make those setups useful.
It also found the sentence in the About page saying readers don’t need to be engineers. Its qualification was fair: following some of the setup articles still means getting comfortable with files, configuration, and technical tools.
That distinction matters here. A title like The $30 Hermes Stack or How I Built a Substack API With Hermes and Codex could make the archive look intimidating to someone who doesn’t code. The agent found enough surrounding explanation to understand who those articles are trying to help.
It then chose these three posts, in this order:
First, How to Build AI Workflows When You’re Tired of Optimizing Prompts. Its reason was the familiar starting point: copying between chats, repeating context, and holding the whole process together yourself. That gives a newcomer a reason to care about workflows before they start configuring one.
Next, Tear Down Your AI Workflow and Rebuild It Like This. The agent picked out the lesson about checking whether your sources serve the work before connecting more services. It’s a useful next step once you’ve decided what you want the workflow to do.
Then, Why My Best Hermes Workflow Consists Of Plain MD Files. This brings the reader to the files themselves and how an agent finds the current document among everything else you’ve saved.
The titles and article text were checked through search after the answer came back. All three were real, and the reasons fit the reader in the prompt.
I wouldn’t call that the only good route through AAC. Someone trying to reclaim their mornings might start with the morning briefing. Someone worried about losing access to a provider might go straight to the vendor-lock-in piece.
What made this answer useful was that the agent explained its choices. I could see how each recommendation followed from the person I’d described.
Where It Got Stuck
The odd detail was that the agent couldn’t directly open any of the three articles it recommended.
Its browsing tool returned “Cache miss” for those requests. The homepage and archive opened, but the article text came through search results instead.
Search had saved enough of the pages to return substantial text. That let the agent read about the workflows and make recommendations even though opening the pages directly had failed.
This distinction is easy to miss when the final answer looks complete. A linked title doesn’t tell you how the agent obtained the information behind it.
In this run, search gave it enough evidence to explain the articles. It couldn’t use that same evidence to confirm exactly what someone would see on the live pages, including whether any content was behind a paywall. Some of the search copies were weeks old.
The same problem showed up when I asked about subscribing.
The agent found public descriptions of Hermes 101, community chat, and direct support. The version of the About page returned by search marked those items “soon.” A later check of the indexed page found the same wording.
So the agent treated them as announcements. It couldn’t verify an enrollment page or confirm that someone could access those benefits that day. The subscription page also failed to open, leaving current paid prices and benefits unresolved.
Its answer on the course was short:
Announced/in development, not verified available.
That was a reasonable place to stop. An older announcement can’t establish whether something launched yesterday, and a failed page request can’t settle it either.
It was careful with the free subscription too. It found an older invitation to receive new posts for free, but didn’t assume that meant every article, course, or resource in the archive was included.
What This Tells Me
The original idea for this article was to find out what an outside agent got wrong about AAC. This run didn’t give me that story. It understood the publication and made useful recommendations.
It did show where its answer became incomplete, and why.
An agent recommending something has to do more than recognize the topic. If it’s helping someone decide whether to pay, it also needs current information about what they’re buying. My archive gave this agent plenty to say about the writing. The sources it could retrieve left it less certain about the offer.
That doesn’t prove my website is broken. It shows what this particular agent could confirm through its tools on that run. Another attempt might reach pages this one couldn’t.
The practical lesson for me is to ask for that detail. If I’d only requested a flattering two-sentence description, I would have missed the part where the agent had to stop.
I covered a related habit in the twenty-minute workspace audit: before blaming the answer, look at the material the agent could find. Here, that meant following its public links and reading its access notes.
For your own work, an awkward answer is worth examining. It might contain an invented claim. It might be working from an old page. Or it might be a fair reading of copy that doesn’t say what you thought it said. Each calls for a different response.
Send this to someone whose newsletter, consulting site, or public project could use a fresh pair of eyes.
Try It on Your Work
You can run the same experiment on your newsletter, consulting site, or public project. Start a fresh conversation with browsing available and leave out the private explanation you usually give your agent.
Choose a specific reader or customer. Someone considering your consulting service needs different answers from someone looking for their first article on a subject.
Then paste this:
I am considering subscribing to, hiring, or using:
[PUBLIC URL]
I am [DESCRIBE THE READER OR CUSTOMER AND WHAT THEY NEED].
Use public sources only. Do not sign in or use private files.
Explain what this person, publication, or project does.
Who is it for, and what knowledge would I need to get started?
Recommend three existing pages or articles to read first.
Give their exact titles, links, and why each fits my needs.
What can I read, buy, or use today?
Separate available offers from announcements and future plans.
Who should skip this?
Cite your sources. Say when you are interpreting something.
If you cannot verify a claim, say so.
Finish by listing which pages you opened, which information
came from search results, and what you couldn't access.
Save the first answer before correcting it. Open the recommended links and check whether they support what the agent said. Pay particular attention to anything involving availability, prices, or promises to a customer.
If a description feels wrong, look for the page that led to it. An outdated sentence is a concrete thing you can fix. An access failure needs further investigation. Neither requires rewriting your whole site on the spot.
Keep any public changes under your review, as with the approval gates I use for agent actions. The agent can help identify a confusing sentence. You still decide what your business promises.
This is one reading of your work. It won’t tell you how every AI describes you or whether changing a page will bring more subscribers. It will give you a description to check against what a stranger can find.
If you try it, what did the agent understand, and what couldn’t it verify?
Top comments (0)