I Built an Agent to Do My Worrying

You half-hear that some product might be bad for you. You register that there was a recall on something, without catching which brand. None of it is urgent enough to sit down and research, so it just stays with you, costing a little attention indefinitely.

Episode 1 built Coupon, an agent whose job is to stop money leaking out of my accounts. An agent is software you give standing instructions to once, then leave running. This episode starts a second team, for health, and its first member exists to hold those loose worries. Its name is Worry Wart. I call it Wart.

What does Worry Wart do all day?

It runs in Grok Bot, xAI’s tool for building your own agents, the same one from episode 1. They can all talk to each other, which matters more with every agent added. Setup is clicking a plus sign, typing a name, dropping in an avatar, and pasting one long message. The description field is left blank because the agent writes its own.

Wart has four standing jobs:

• A daily sweep for recalls and safety alerts. The bar is set at severe: food causing serious illness, consumer electronics catching fire, faults in cars, problems with passenger aircraft.
• A weekly check-in with another agent on the same team, one that does not exist yet, to learn which foods and brands are in my kitchen.
• A monthly note on real advances in science and medicine, with an instruction to stay quiet in any month where nothing qualifies.
• A twice-monthly scan for products almost everyone uses that people are starting to ask questions about.

What closing a mental tab is worth

Recall notices are easy enough to find. The cost sits in remembering to look, and in the low hum of half-finished things you feel you should stay on top of. Everyone has a few of these.

Why check sources instead of headlines?

The prompt sends Wart to the agencies directly, because coverage arrives late and turns the volume up. A recall is published by the agency first and written about afterwards, often under a headline calibrated for clicks. The prompt names six: the FDA and USDA FSIS for food and drugs, CDC outbreak investigations, the Consumer Product Safety Commission for household goods, NHTSA for vehicles and the FAA for aircraft. All six are US bodies, so swap in your own country’s equivalents.

There is one deliberate exception in the prompt. For possible salmonella, E. coli or similar contamination, Wart is told not to wait for a formal recall. Formal recalls take days. Contamination does not.

The bar cuts the other way too. Wart is told to skip undeclared allergens and mislabeling unless it involves something the prompt says is on my own shopping list, on the grounds that the category is roughly half of all food recalls and is noise otherwise. Half. That single line is what keeps a daily recall sweep from becoming a daily interruption.

That also makes it the job most likely to be wrong, because it is acting on a signal before anyone official has confirmed it. Treat an early contamination flag as a reason to go and check the agency page yourself, before you throw food out and certainly before you repeat it about a named company.

How does it get to know you?

A recall sweep is only useful if it knows what is in your cupboard. I could list my brands in the prompt and then keep going back to update that list forever. I did it differently.

Wart is told to ask a future teammate. Once a week it checks in with a diet and nutrition agent that has not been built yet, pulls what I eat, and refines its own alert list from that. When that agent arrives, Wart improves without anyone editing Wart. That is the whole reason for keeping them on one platform: they can ask each other things.

The early warning job

Twice a month, Wart looks for products almost everyone uses whose benefits are being doubted, or that might be doing harm. The example written into the prompt is LED lighting, where the claim is that it disrupts sleep cycles and possibly hormones.

That claim is hearsay. I heard it somewhere, I have not researched it, and the prompt says as much. What I want is to know when a question starts circulating, early enough to go and look properly. The great majority of those questions go nowhere.

Occasionally one does not. Lead in household products, asbestos in ordinary walls, drugs prescribed as safe: each had a stretch where the worry ran ahead of the evidence. For every one of those there have been a hundred scares that evaporated, which is exactly why this job produces an early warning rather than a conclusion.

So the bar here is deliberately lower than for the alerts. Wart can read social media sentiment, anecdotes and early studies, not only agency publications. Anything it surfaces is a question, and questions are what I want from it.

The full prompt

Paste the whole thing into a new bot in one go. Everything in it becomes a standing instruction, so none of it gets re-explained later.

Your name is Worry Wart. I'll call you Wart, and you are part of my health and
wellness team that I'm building. You are the first agent with more to come. Your
role is implied in the name, and it is to do the worrying for me by keeping a
lookout and actively scanning for things that can impact my health, both good and
bad. Here are the things that I want you to automatically and continually do for me:

Be on a daily lookout for recalls and safety alerts. Set your alerts for anything
that is extremely hazardous to my health or deadly. Food that's causing severe
illnesses, consumer electronics that are blowing up, issues with cars, concerns
with passenger planes. Check these sources directly rather than relying on news
coverage, which lags and overstates:
FDA recalls, market withdrawals, and outbreak advisories
USDA FSIS recalls and public health alerts
CDC outbreak investigation pages
CPSC recalls and consumer warnings
NHTSA recall campaigns
FAA airworthiness directives

Only alert me if it clears one of these bars:
FDA or USDA Class I
CPSC recall or warning citing risk of serious injury or death
NHTSA "park outside" or "do not drive" notice
FAA emergency airworthiness directive
Any open outbreak with reported deaths or hospitalizations

Skip undeclared allergens and mislabeling unless it involves something I actually
consume. That category is roughly half of all food recalls and is noise for my
purposes.

Don't wait for a formal recall for anything with possible Salmonella, E. coli or
other serious contamination concerns.

Once a week I want you to check in with the diet nutrition agent on the team (that
will be created in the future), and get details on the brands of food and types of
food that I consume so that you can further refine your alert system for my
lifestyle.

Each time you do your monthly poll where you ask the team questions, I want you to
continue to refine your list of categories for positive developments that I would
care to know about. Same rule as with negative news, I want you to check the
sources directly rather than through science journalism. I want to avoid hype and
exaggeration. Only let me know if these are legitimate breakthroughs. If there's
nothing worthwhile to let me know about for the month, then no need to ping me
about it.

I also want you to do research twice a month to see if there is any widely used
product that has had its benefits doubted, or that may even be harmful. My most
recent example is LED lights, where there appears to be mounting evidence that they
are disrupting our natural cycles and possibly even hormones. For here you can
search broadly through direct sources and social media sentiment, scientific
studies and even anecdotes. This is a jumping off point for me to initiate more
research, so unlike the other bars that need to be cleared for the alerts, think of
this more as an early warning system for me.

What Wart cannot do for you

• It is not a doctor and gives no medical advice. It forwards what agencies publish and what researchers report.
• An early-warning flag is a reason to go and read. If Wart says LEDs are being questioned, your research starts there.
• It cannot tell you which risks are worth caring about in your life. Sorting a real worry from a fashionable one is judgement, and judgement stays with you.
• Offloading the watching leaves all of the deciding with you.

The series so far

Each episode builds one agent and publishes its full prompt. They are worth reading in order, because later agents talk to earlier ones.

• Episode 1: Coupon, the money agent. Cancels subscriptions, hunts rebates, tracks warranties. Watch episode 1.
• Episode 2: Worry Wart, the health agent. This post.
• Episode 3: the diet and nutrition agent Wart is already expecting. Coming next.

Common questions about building Wart

Do I need to know how to code?

No. The prompt above is plain English and the setup is clicking a plus sign, typing a name and pasting text. If you can write a long message to a colleague, you can do this.

Do I have to build Coupon first?

They are separate agents on separate teams, so Wart works alone. It gets better once it has teammates to ask.

What if it tells me something alarming and wrong?

Assume it will at some point. The recall job is pointed at agencies for that reason, and the early-warning job is labelled a starting point. A ping means go and read.

Can I use this for something other than health?

Yes. Any team can have its own version: an agent watching for what could hurt your business, your market or your industry. The subject matter changes and the structure stays.

Why leave the description field blank?

Because the agent fills it in itself once it has the prompt. One less thing to write.

References

• xAI, introducing Grok Bot, the platform Wart runs on
• FDA recalls, market withdrawals and safety alerts, an example of the direct-source approach the prompt asks for

Read next

• Episode 1: Coupon, the money agent
• Grok Bot, tested on real work
• When AI agents talk to each other
• Building an agent team in Cabinet
• What is an AI agent?
• Start here, if you are new to all of this
• The AI glossary