Practitioner's Corner

Practitioner's Corner




One Diagnosis and Three Hundred Permissions, an Interview from the Agent Help Queue
CONTINUE READINGBrief Counterpoint

Further Reading


© TinyFish, Inc.. All rights reserved.





Brief Counterpoint

Further Reading


A grocery store's computer says the shelf holds six cans of tomatoes. The shelf is empty. Because the computer believes the cans are there, it never orders more, and the shelf stays empty while the records say all is well.
The same blind spot, a check that waits for events the failure has stopped, shows up in how OpenAI monitors its AI agents. It may also show up in how you monitor yours.
A grocery store's computer says the shelf holds six cans of tomatoes. The shelf is empty. Because the computer believes the cans are there, it never orders more, and the shelf stays empty while the records say all is well.
The same blind spot, a check that waits for events the failure has stopped, shows up in how OpenAI monitors its AI agents. It may also show up in how you monitor yours.
The code in curl runs in billions of devices, so when someone reports a security flaw in it, its maintainer has to find out whether the flaw is real. AI tools have made such reports cheap to send, but each one is still slow to check. By May, more than one report a day was arriving, nearly all of them AI-assisted, about twice the 2025 rate. This profile follows a maintainer under that pile who is betting that strangers will step in to help.
The code in curl runs in billions of devices, so when someone reports a security flaw in it, its maintainer has to find out whether the flaw is real. AI tools have made such reports cheap to send, but each one is still slow to check. By May, more than one report a day was arriving, nearly all of them AI-assisted, about twice the 2025 rate. This profile follows a maintainer under that pile who is betting that strangers will step in to help.
In a Johns Hopkins trial published in 2024, young people with diabetes needed an eye check, since diabetes can damage the eyes. Families handed a guide and told to book an exam elsewhere mostly didn't. Families whose child was photographed during the diabetes visit they had already come to, with AI reading the images on the spot, all completed it.
Then came the step nobody redesigned. A child whose scan suggested damage still needed a separate specialist appointment, booked by the family. About a third hadn't made it six months later.
For anyone automating a workflow, both halves matter. A handoff is a point where people must go elsewhere and arrange the next step themselves. Removing one produced a large, real gain. The losses then gathered at the next phone call someone had to make. And because the software and the same-room exam arrived together, the trial can't say which did the work.
In a Johns Hopkins trial published in 2024, young people with diabetes needed an eye check, since diabetes can damage the eyes. Families handed a guide and told to book an exam elsewhere mostly didn't. Families whose child was photographed during the diabetes visit they had already come to, with AI reading the images on the spot, all completed it.
Then came the step nobody redesigned. A child whose scan suggested damage still needed a separate specialist appointment, booked by the family. About a third hadn't made it six months later.
For anyone automating a workflow, both halves matter. A handoff is a point where people must go elsewhere and arrange the next step themselves. Removing one produced a large, real gain. The losses then gathered at the next phone call someone had to make. And because the software and the same-room exam arrived together, the trial can't say which did the work.
Booked elsewhere: 18 of 82 children whose families got a guide and a referral were screened within six months.
Screened at the visit: 81 of 81 analyzed children offered the in-clinic AI exam were screened.
Flagged by the AI: 25 of those 81 screened children got a result suggesting eye damage. This was the software's reading, not a diagnosis.
Reached a specialist: 16 of those 25 flagged children saw an eye doctor within six months. The trial had no comparison group at this step.
Human backup: A specialist also reviewed every photo, because the AI wasn't cleared for patients this young.
Not separated: The AI reading and the exam happening in the same room arrived as one package.
Not measured: Treatment completed, vision protected, staff and family time spent.
Who: 164 participants aged 8–21, enrolled November 2021 to June 2022.
Past Articles

On October 21, PHP, the language running roughly seven in ten websites whose server language can be identified, shipped ...

When a person takes control from an AI agent driving their web browser, the system records a human step, but not why. Fa...

Ananya Gubbi Mohanbabu, a UC Berkeley PhD student, builds tools for blind and low-vision people who use screen readers, ...

Boston gave residents seven categories for reporting city problems, plus "other." When a researcher read a sample of rep...