AI agent sends personalized Mid-Autumn greetings to 50 contacts by seeing the screen and clicking, then replies to 70+ messages
FuSheng_0306 · x · 2026-09-27
A Chinese author ran a Mid-Autumn experiment letting an AI personal assistant send personalized greetings to 50 colleagues and reply to 70+ incoming messages — with no APIs, scripts, or backend access.
Key details
- Personal WeChat has no open API, chats are encrypted, and system-level screenshots are blocked, so batch-scripting was impossible.
- The agent converged on a human-like path: visually read the screen to find the contact, chat window, and input box, then control mouse and keyboard step by step — looking to the program like "a person operating the computer."
- For each greeting, the AI asked the author who the person was, what they'd done lately, and what he expected; he dictated fragments ("he's been carrying a project alone"), the AI polished them, and the author confirmed before sending.
- The AI also replied to incoming messages, adapting tone based on the relationship and context.
- He contrasts this with a scripted Spring Festival blast to 611 people, which was mere automation; the new capability on display is long-horizon, tedious, context-aware task execution over thousands of uninterrupted steps.
- Boundaries: no spam, no API exploits, no chat-log scraping, no template flooding — human confirmation for anything important.
More from coding & agent
- Dev argues coding tools don't need a dedicated plan mode: keep humans in the loop always — jonathan_wilke · 2026-09-27
- openrig: multi-agent harness running Claude Code and Codex together as one system — mvschwarz · 2026-09-27
- Vercel open-sources scriptc, a TypeScript-to-native compiler — vercel-labs · 2026-09-27
- Open-source Claude Code skill /brag turns any repo into a launch video with one command — petewoodbridge · 2026-09-27
- Adding memory to an incident-response agent: Hindsight's reflect() caught two forgotten outages — Rajitha29 · 2026-09-27
- Managing 5 LLM providers: which gateway handles billing, spend caps, and fallbacks? — InsuranceDependent71 · 2026-09-27