Decide what it can touch.

General-purpose AI agents like ChatGPT Work, ChatGPT Codex, Claude Cowork, and Claude Code can now work through files and apps on your own computer. You give one a request, it picks its own steps, and it uses software tools to get the job done: find a document, update a spreadsheet, draft a message for you to review.

A new comment in The Lancet Digital Health from Mahmud Omar, Eyal Klang, and Girish Nadkarni (Mount Sinai and Beth Israel Deaconess) asks the obvious healthcare question. How do you keep an agent inside the clinical task it was given?

At its Open@Epic developer conference on Oct. 21-22, Epic said outside developers will be able to build Connectors that link their software to Agent Factory, its platform for building and running AI agents. Early-adopter customers could start as soon as next month, and developers can start building in the first quarter of next year. Epic also said consumer chatbots are adding real traffic to health systems' records. And a Costco pharmacy leader said its pharmacies get about 1,000 calls a day from AI agents asking about controlled substances.

Researchers using frontier AI models found serious flaws in PHP, the software behind a huge share of the web, including WordPress, Drupal, and Joomla sites. Security firm Wiz showed an unpatched site can be taken over remotely with no login, and that attackers can read passwords and keys straight from the server's memory. Wiz says every PHP version back to 5.4 is affected. The PHP project shipped emergency patches on Oct. 21.

A team at the Medical University of Warsaw had three arthroplasty surgeons rate ChatGPT's answers to 70 patient questions about total knee replacement, including 20 about fast-track recovery (early walking, early discharge, fewer opioids, fewer drains and catheters). Average scores sat near 5 out of 5 for accuracy, completeness, safety, and readability. Every fast-track answer got a perfect score from every rater.

Bar-Ilan University in Israel piloted an AI agent on its Moodle site across 71 courses and about 2,400 students. It answers only from material faculty uploaded: recorded lectures, transcripts, slides, and articles. Students can search it, get summaries, and quiz themselves, and it's expanding to 100 more courses this fall. One lecturer spent real time rewriting and expanding his materials, then tested the answers before he trusted it. His class average went up three points, which he admits could just be a stronger group.

A team at Charité Berlin ran 122 expert-scored summary statements from virtual patient cases through four models (Claude, GPT-4o, Mistral Large, and Mistral 7B), five times each, under four prompting styles. The three large models gave nearly the same scores from run to run. Agreement with the human experts was best when the prompt included a few graded examples and asked for step-by-step reasoning, topping out at 77% accuracy for Mistral Large and 71% for Claude. The models did better on yes-or-no items like factual accuracy than on the three-level judgments about reasoning quality.

Before you let a desktop agent loose on a clinic computer, make it a folder of its own. Put in only the files it needs for one job, with no patient records and no shared drive. Point the agent at that folder and nothing else, and turn on whatever setting makes it ask before it sends, deletes, or changes anything. Then run one real task, like cleaning up next month's staff schedule, and watch what it reaches for. If it asks for more access than the job needs, you've learned something about the tool before it learned anything about your patients.