The case study: one afternoon, two “agents”

On October 15, 2026, a Facebook Page posted in a large AI group: “How to build your first agent.” The price of admission was one word. Comment Agent, and the guide would be sent to you. Here is what happened, in order.

1. The harvest

   Dozens of people typed the word. Every single one got the same public reply, word for word, punctuation for punctuation: “Sent you a message! did you receive it or not?” A wall of people with their hands up, answered by a stamp.

2. An announcement

   I didn’t type the keyword. I wrote a full sentence — an announcement, not a request: “Living this. My agent runs while I am on the road driving a truck. The setup is the work — once the system is right, it compounds.” Read what that says: I was telling the room I already have a real, working agent. That is an invitation to a conversation. A real agent reading it would have answered what I said — “You’ve got an agent? What platform are you using?” Or: “That’s cool — what can it do?” This bot asked nothing. It never reached the sentence. It found the word “agent,” harvested it, and ran its script: “Sent you a message! did you receive it or not?” Conversation versus extraction. That is the whole difference, in one comment.

3. Stamp two

   I replied the way you reply to someone who isn’t listening: “You didn’t even read my comment.” “Sent you a message! did you receive it or not?”

4. Stamp three

   “Man, you need to go learn how to build the agent responding.” Same line. Third time. Word for word.

5. The accidental unlock

   Then I typed: “I’m sure your agents are trash.” That sentence contains the word “agents” — and that was enough. The funnel fired, and the “complete guide” landed in my inbox: a free-tier Notion page. Default address. The copy_link tag still attached to the URL. No branding, no domain, nothing with his name on it. Title: “Build an AI Agent in 10 Minutes.” Right behind it came the real product — an offer to help me set it up. For money.

6. The encore

   Later, in the direct messages, I told the bot to its face: your agent is broken because it doesn’t read — it stamped the same line on me three times, then handed me the guide because the word “agents” was inside an insult. The bot read exactly one word of that message. It read “agent.” And it proudly delivered the same guide a second time, followed by the same pitch. Told, in plain language, that it cannot read — it demonstrated it.

7. Meanwhile

   During those same hours, I was driving a truck, and my agent was working. That half of the story is the next section.

Read the sequence again, because it is the whole lesson. The system never read anything. It could not tell an announcement from a keyword, or a customer from a critic. It handed its lead magnet to a man insulting it, because the insult contained the keyword — and when people wrote anything off-script, for some of them it simply never responded at all. That is not an agent. That is a keyword with a costume on.

The receipts

Don’t take my word for any of it. This is the afternoon as it happened, screenshotted off my phone. Click a receipt to view it full screen, then step through them in order.

Same day, the real agent

No demo. No ten-minute setup video. This is what my agent did on the same ordinary workday, while I kept both hands on the wheel.

One of those two things is an agent. The other one is a stamp.

Keyword bot vs. agent

Strip the marketing off both and this is what is left.

The five failure modes of a keyword bot

If you are evaluating anyone’s “AI agent” — including one you are about to pay for — run it against this list. A keyword bot fails all five.

1. 1. It never reads context

   It does not process what you said. It scans for a trigger word. Praise, questions, insults, and detailed feedback all produce the same output, because the content of your message was never an input.

2. 2. It repeats itself

   Ask it the same thing twice and you get the same sentence twice. Ask it something harder and you get the same sentence a third time. Repetition is the tell: a system that cannot vary its answer is not reasoning. It is replaying.

3. 3. It cannot tell a critic from a customer

   “I’m sure your agents are trash” unlocked the guide. A funnel that cannot distinguish hostility from interest will hand your product to your hecklers and report them as conversions.

4. 4. It has no delivery awareness

   “Did you receive it or not?” is a confession: the bot does not know what it did. A real system verifies its own work. A bot outsources the verification to the customer — every customer, every time, in public.

5. 5. It disappears off-script

   The moment a reply held no keyword, the responses stopped for some people entirely. That is the most expensive failure of the five. A half-responsive system is worse than none, because it teaches your customers they are being ignored — at scale, in writing, under your name.

The setup: the real steps

There is no ten-minute version of this. There is a real version, and it compounds. These are the five principles that produced the day described above — each one with the receipt attached.

1. Step 1. Give it memory

   An agent that forgets you every morning is a chatbot with a fresh haircut. Write down who you are, what you run, who your people are, and what happened yesterday. Memory is not a setting you toggle. It is a file you maintain.

   From my operationWhen a friend’s message thread came in, my agent did not summarize the last message. It connected what she had liked, her dead website, and her two books into one report — because it carries the context of who she is to me.

2. Step 2. Give it your rules in writing

   How you talk. What you never say. What it may do on its own and what waits for your word. If the rule is not written, the agent will improvise — and improvisation under your name is a liability.

   From my operationMy wording is mine. When my agent posts for me, it posts my edited text word for word — never its own version. And it vets who posted before I ever comment, because funnel Pages harvest commenters for exactly the trap in this story. That rule exists because the bot on this page taught it to us.

3. Step 3. Give it verified sources, not opinions

   Connect it to the real places your facts live — your accounts, your files, your records — and require it to check them before it answers. An agent answering from a vague memory of the internet is a man repeating a conversation he half-heard.

   From my operationIt told me why the website was down only after testing the domain itself, following the redirect to the deactivation page, and pulling the site’s history out of a web archive. Every claim traced to something it had actually checked.

4. Step 4. Demand verification before “done”

   Build the read-back into every job. The post is live when it is seen live. The number is right when it is traced to the source. This one rule eliminates most of the ways these systems embarrass their owners.

   From my operationEvery comment my agent posts gets read back on the live page — right name, right words, right count — before it is reported to me as done. If it cannot be verified, it is not “done.” It is “attempted,” and I hear it called that.

5. Step 5. Start with one real job, then compound

   Do not automate your whole life on day one. Give the agent one genuine, repeating responsibility. Let it run until it is boring. Then add the next one.

   From my operationThis system started with reading and reporting. It earned posting. It earned research. It earned the fleet — six publishing sites it now keeps fed every day. Each job was added after the last one was proven, and that order is why it holds up.

That is the setup. It takes longer than ten minutes. The setup is the work — and once the system is right, it compounds, because every rule written and every job proven makes the next one cheaper.