AI Agents Are Running Errands Now
I used Claude Cowork for the first time, and it blew me away. I asked it to check on an at-home radon test I’d mailed off and forgotten about. About ten minutes later, an inquiry email had gone out from my Gmail, and I hadn’t opened a single tab or picked up the phone.
The errand
Six weeks earlier, I’d mailed off an at-home radon test kit and promptly forgotten about it. So I asked Claude to check on the results. Beyond that one request, I didn’t tell it how. It worked out the rest:
- It found the testing company’s website.
- It discovered that the “results lookup” page was a dead end.
- It dug through the company’s documentation and confirmed that results are only delivered by phone or email.
- It came back with both options and a drafted inquiry, ready to go.
I told it to send the email from my Gmail, with a BCC to the alias I use to track threads I’m waiting on. It opened my browser, composed the message, hit a few formatting snags along the way, caught and fixed its own mistakes, verified everything was correct, and sent it.
The whole thing took about ten minutes, and I spent them on other, more important things.
The dead end was the interesting part
A script would have stopped at the results lookup page. Plenty of people would too: you find the page that says “results,” it doesn’t have yours, and the errand goes back on the list.
Claude treated the dead end as information. It went to the documentation to find out how results actually get delivered and came back with the real answer: phone or email, and nothing else. Noticing that the obvious path goes nowhere and finding the one that does is the part of an errand that usually needs a person.
Then it handed the decision back to me. It laid out both options with a draft ready, and I chose email and told it to send. The agent did the research and the drafting, and the call to send something under my name stayed with me, which is the split I want.
It checked its own work
It wasn’t a clean run. It hit a few formatting snags along the way, caught them itself, fixed them, and verified everything was correct before it sent anything.
That’s the same pattern that makes coding agents useful. They make mistakes, and that’s fine as long as they find out fast and correct course before building more on top. Composing an email in a browser has a feedback loop too: look at what’s on the screen, compare it to what you meant, fix the difference. Sending an email can’t be undone, so the check before that step is the one I care about most.
Handing an agent the keys
This errand worked because the agent could act as me, in my browser and my Gmail. That’s also what makes it worth some care.
An agent running errands reads pages written by strangers and then takes actions with your accounts. Any text it reads is a place someone could plant instructions. I’ve seen how well that works from the other side: a prompt injection in my LinkedIn profile was enough to steer the recruiting bots that read it, which I wrote up in My AI Recruiter Honeypot Worked.
A testing company’s documentation is a low-risk place to send an agent. An inbox full of unknown senders is a different story. The habit from this errand is the one I’d keep everywhere: let the agent research and draft freely, and keep the decision to send something in my name with me.
Call it a time maker
Ten minutes of agent time is a nice saving. The bigger change is that “check on my radon test” went from a vague item on my mental to-do list to a completed task.
Errands like this are short and annoying. Each one needs a small block of attention: find the right page, discover it’s useless, dig up a phone number or an email address, write the message. This one had been sitting, forgotten, for six weeks. Claude finished it while I was doing something else.
So call it a time maker. It made time for a task I hadn’t made time for myself.
AI agents are running errands now.