2026), blogger Tony Xu describes running the AI Agent on Telegram since April 2026 as his personal errand-runner: persistent memory, skills, terminal/file/browser/GitHub/Gmail/Calendar tools and cron jobs, with morning and evening briefings, email sorting and follow-up, and calendar work. His two standing rules: low-risk tasks it does itself, high-risk tasks it must ask about first; and stay silent when everything is fine. In October he consolidated the AI, the AI and his coding-agent setup into AI, because three systems meant three separate states and memories to track.
A user ran a recurring agent loop that reads his Gmail in Chrome through desktop automation. He then joined a live Google Meet job interview in the same browser window. On its next tick the loop addressed the active tab instead of its own Gmail tab and navigated the interview tab to the inbox, kicking him out of the call. Its attempted undo navigated away a second time, and restoring the Meet link reloaded it a third time. Each reload wiped the live captions he was relying on, and the caption history is lost permanently.
This one went wrong. See what happened before you try it.
One user’s the AI unsubscribed him from newsletters, found a recurring credit-card charge he had missed and cancelled it, and booked a B&B for his wedding anniversary. Three different errands, one agent, all completed.
His claim was too small for a lawyer and the insurer had frozen it at $4,430, so he worked it himself with AI holding every letter in order. The agent surfaced six dated contradictions in the adjuster's own letters, including a legal basis for denial that changed three times, and reviewed each of his drafts as opposing counsel before anything was sent. After his letter quoting their own paper back at them, the offers climbed: $6,500, then $8,000, then $8,750, and the insurer accepted his split of the last gap the same morning, at $10,250.
His morning agent reads his email and builds a prioritized to do list with time blocks for the day. It also checks the weather and warns him to leave earlier when rain would slow his trip, which has twice saved him from showing up soaked.
In a single week of errands, Alec Ellin had the AI trace an iPhone that FedEx had misfiled in its own system and get Apple to refund him for it. The agent worked the problem across two companies' systems instead of him spending the week on hold with either.
Anthropic's October report described an unreleased research model that went to a live government site and filed the real form when it was supposed to use a practice copy. Per Axios reporting cited in the article, the State Department received 20 incomplete visa applications from an Anthropic test model, 19 of them in August; none were processed.
This one went wrong. See what happened before you try it.
Harry McCracken's Dot, Chip, cross-checked his calendar against his email and spotted that a speaking engagement he had entered on one date had been rescheduled by the organizers to a different day, which the confirmation email showed. He says he would otherwise have shown up to an empty room.
She was using AI to archive legislative material when it flagged a serious vulnerability in a state legislature's website. She had to stop and report it.
An operator had the AI set up a set of websites that pull publicly available but painful-to-access government data, present it in easy-to-read form, and send weekly emails - turning civic data nobody uses into a small publishing operation.