What they did
Hassan Abid, CTO of a small company, had about two weeks to prepare for surgery and time away from both work and household duties. Abid set up Hermes and OpenClaw for home and scheduled coding-agent routines for work, all on one Mac mini. Before the operation, Gemini models read Abid's medical records, both insurance policies, an 87-page plan guide and photographed slides, citing page numbers, and filled in the claim forms. Gemini also wrote scripts for scheduled reminders, which Hermes then ran without calling a model and sent over WhatsApp. After surgery, Hermes kept one Markdown file of blood panels, imaging, medications and blood pressure from phone photos. OpenClaw watched the inboxes and followed the insurance claim. At work, routines treated Slack as context, posted labelled Jira comments, drafted email forwards and ran a read-only cloud alarm check.
How it went
The agent found an exclusion in a spouse's plan and a warning on a photographed slide the author might have missed, and the claim was eventually settled. The recovery tracker lasted after chats were cleared and went to appointments. The email routine drafted nine forwards and sent none, as intended. The cloud check reported no active alarms for weeks while one alarm fired 318 times in two weeks; the script split alarm names at spaces and missed 20 alarms. One routine sent a Slack message from Abid's account that a colleague took as sent by Abid in person. A timed reminder told Abid's wife surgery was underway based only on the schedule.
Worth knowing
Abid found the guardrails worked better than the monitoring. The fixes the author lists: label automated messages clearly, write down who can decide in your place, verify that a check actually covered every item before trusting an all-clear, and word time-based reminders as plans. The couple still asked the insurer for written pre-authorisation rather than relying on the agent's reading, and notes how much private data went to a hosted model.