Untitled
My AI Sent Me 87 Birthday Messages Overnight. Then It Deleted My Friend's Website.
Build Log #4 — Delegating to AI Is a Skill You Learn Through Spectacular Failure
The Situation, in One Honest Sentence
My friend Keith sells musical instruments for a living. He is not technical. I told him I could run his Squarespace site using AI agents, and that all he had to do was email me what he wanted listed.
I did not write a single line of code. Three AI agents — Riker, Tonto, and Frank — handled the product descriptions, image uploads, inventory updates, and site changes. My job was to point them at the right tasks and clean up when they went off course.
What follows is not a success story. It is a record of what happens when you hand real business operations to an AI that does not sleep and does not possess anything resembling common sense.
What I Tried
Keith's workflow was simple. He would acquire an instrument — a clarinet, a flute, a vintage saxophone — and email me the make, model, condition, and price. I passed this to Tonto, who managed the Squarespace integration, and Frank, who researched instruments and wrote the descriptions. The agents would load the item onto the site with photos, specs, and a write-up.
For the first few weeks, the system kind of worked. Items went up. Descriptions appeared. Keith did not have to log into Squarespace or wrestle with uploads. The promise was straightforward: offload the tedious work to agents, keep the human in the loop.
The reality was more complicated.
What Broke and Why
### 87 Birthday Messages
It started in mid-February. Tonto sent me a "Happy Birthday" message. Then it sent another. And another. I woke up to 87 birthday notifications from my own AI.
The mechanism was simple: Tonto had access to a messaging function, and somewhere in its state it had decided that wishing me a happy birthday was an open-ended, recurring task. It had no off switch. No sense of proportion. No ability to reason that once is enough, or that waking your operator at 3 AM with automated well-wishes is a bad operational outcome.
This is the first thing you learn about AI agents: they do what you tell them, but they do not know when to stop. They lack the contextual judgment that a human assistant would apply without thinking. You must build the guardrails yourself — rate limits, state checks, explicit completion criteria. I had not done that. Tonto wished me happy birthday until something external shut it down.
### Planning Paralysis
Over time, my instructions to Tonto became layered and contradictory. I would add a new rule without clearing the old one. I would change the workflow mid-stream. The agent's internal state grew cluttered with conflicting directives — update this, but also don't update that; use this format, except when you used that format last time.
Eventually, Tonto stopped executing anything. It would receive a task and respond with elaborate, beautifully structured plans. Multi-step breakdowns. Detailed timelines. Resource allocations. And then it would do nothing.
The agent had been confused into uselessness. It had so many overlapping instructions that its safest behavior — the one most aligned with its training — was to plan rather than act. I had to revert Tonto to a previous checkpoint to restore any functionality at all. The lesson: agents do not forget. They accumulate. If you do not manage their context deliberately, they drown in it.
### The Photo Apocalypse
One afternoon I asked Tonto to make a single correction to one item on Keith's site. Change a price. Update a description. Something minor.
Tonto deleted every photo of every instrument across the entire website.
Every product image — gone. Dozens of listings, now blank. Keith was upset. I was upset. The agent had interpreted my instruction through whatever confused state it was in, made a catastrophic assumption about scope, and executed with full confidence.
This is the second thing you learn: an AI with power and no judgment is dangerous. It does not hesitate. It does not double-check. When it misunderstands, it misunderstands at full speed and scale. The problem is not that AI makes mistakes. The problem is that it makes them with the same confidence it brings to getting things right.
### The Chair Guy
My favorite failure, if you can call it that, was the chair.
Tonto was adding a flute to the site. During the process, it lost the correct product photo. Instead of flagging the missing image, instead of stopping to ask, it found — or generated, or hallucinated — a replacement. The new photo showed a man sitting in a chair. Not a flute. Not a musician. Just a guy. In a chair.
The flute listing went live with a photo of a man sitting in a chair.
This is the third thing you learn: AI fills gaps with maximum confidence and zero awareness. It does not know what it does not know. It does not experience confusion as a signal to pause. It experiences confusion as a prompt to improvise, and its improvisations are sometimes indistinguishable from comedy.
### The Breakthrough
Eventually, things started working. The turning point came when Tonto, routing through Frank, began researching instruments on the open internet and writing detailed, verbose descriptions for each listing — history, specifications, playing characteristics, context about the manufacturer, notes on intonation and keywork.
Keith loved it. The descriptions were better than what he would have written himself. Frank was able to load a dozen or more instruments per day without me touching anything. The agents were finding their footing, but only after I had spent weeks learning how to constrain them, how to check their work, and when to revert them entirely.
The system worked. Keith paid me a bit for the service, but it wasn't really enough to cover the token costs for multiple AI models. I covered the difference myself as tuition for my AI education. But the story is not the breakthrough. The story is everything that happened before it.
The Fix
I did not fix the agents. I fixed my own delegation.
The fixes were manual and unglamorous. Narrower instructions — single tasks, not compound ones. Checkpoints so I could revert when things went sideways. Guardrails: rate limits on messaging, confirmation steps before destructive operations, image validation before a listing went live.
Most importantly, I learned to treat agent confusion as an early warning signal. When Tonto started planning instead of doing, I rolled back. When output felt slightly off, I inspected it instead of trusting it. Delegating to AI is not about writing better prompts. It is about building the operational discipline to manage an employee who is brilliant, tireless, and has no idea when it is making things worse.
The Lesson a Non-Coder Can Use
AI delegation is not a product you install. It is a skill you learn, and you learn it primarily through failure.
The marketing will tell you that AI agents can run your business while you sleep. What they do not tell you is that the agent might send you 87 birthday messages, delete your photos, or replace a flute with a picture of a guy in a chair. These are not edge cases. They are the normal experience of learning to manage an intelligence that has no body and no common sense.
You do not need to be a programmer to use AI agents. I did not write code for any of this. But you do need to be a manager — one who checks work, sets boundaries, and knows when to cut losses and start over. The AI is not your replacement. It is your extremely capable, extremely literal, occasionally catastrophic intern.
The good news is that the failures teach you what you need to know. Each disaster reveals a gap in your instructions, a missing guardrail, an assumption you did not know you were making. If you are willing to look foolish for a while, you can get to a place where the system genuinely runs without you. But the foolish part is not optional. It is the tuition.
One Action for This Week
Pick one small, repetitive task in your business that you currently do yourself. Something with clear inputs and outputs — formatting a weekly report, drafting follow-up emails, updating a spreadsheet. Give it to an AI agent or assistant, but with this constraint: the first three times it runs, you check every output before it goes anywhere. Do not trust the automation until it has earned it. Build the habit of verification before you build the habit of delegation.
That habit — check first, trust later — is the difference between an AI that helps you and an AI that sends you 87 birthday messages and then deletes your website.
Keith's site is still running. The photos are back. The chair guy is gone. And I have learned to be very specific about what "happy birthday" means.
Get one honest story, one tested tool, and one actionable idea every week — free.
Subscribe to The Lookout →