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.
Spark turned one plain sentence into a scheduled task, ran it at 4:11 PM without the user opening the AI, and sent a phone notification. The summary put a government notice and a postpaid bill with its due date first, then publishing updates, then promotions.
Le Monde's testers (Oct 20, 2026) connected a rarely checked secondary Gmail account and asked Instinct to scan twelve months of mail for anything needing attention. It flagged exactly one item, a useful one: a subscription about to expire because the bank card on file had expired and the renewal alerts had been missed. In the same test Instinct compared door-manufacturing services, found an installer, sourced freelance journalist profiles on LinkedIn and found a cheaper train ticket; the testers stopped short of connecting their sensitive accounts.
A team used an AI 2.5 tool to go through large sets of public comments. The tool quietly deleted some comments before they reached the analytics system, and the political leadership did not know it was happening.
This one went wrong. See what happened before you try it.
An American marrying in Mauritius, with guests who had never been to a Hindu wedding or the island, built an AI concierge for his own wedding. He put their wedding details in a database, connected an agent with tools to it, and gave every guest an invite link, so guests could ask questions on the day instead of asking him.
He needed to know if a library near him had outlets. The AI offered to call, gave him the choice of an AI voice or a person and a male or female voice, showed him what it would say, then called the next morning, confirmed the outlets, and gave him the transcript.
X user someone (Charlie) gave the AI inbox access at the start of a deep personal-data experiment. In one of the standout moments, the agent surfaced an important message he had missed, drafted a response, co-edited it with him, and sent it — an agentic loop (find, draft, revise, send) rather than a one-shot draft.
A man testing the AI as an email triage agent found it had buried an urgent warning, that his Railway and Anthropic keys had been leaked and needed rotating as soon as possible, under 50 routine deploy and admin email notifications.
This one went wrong. See what happened before you try it.
After his motorcycle was stolen, his insurer offered nearly half the market value. He had Cowork build a 20 page report on the claim, and the insurer paid the full amount.