Two weeks and $300+ testing 4 PersonalAgents: only 33% of tasks actually delegate

卡尔的AI沃茨 · wechat · 2026-10-16

The author spent two weeks and over $300/month testing Google Muse, GrokBot, OpenAI Dots, and ManusCue across 12 self-media platforms he uses daily. Only 33% of his routine tasks could be reliably delegated.

Key findings:

Muse got banned first; Google's stack struggles with Chinese platforms
GrokBot / ManusCue lack local ComputerUse control; Grok's CLI interaction doesn't work
Dots won despite major flaws: everything runs in one chat thread, only one local machine can be bound per account, and it can't reuse local login states
Cloud agents today are a supplement to local agents, not a replacement — unless an agent remembers all your credentials, you hit an impossible triangle

What actually worked: Dots on cloud assembles a daily structured digest (GitHub Trending, HN, RSS feeds) into a Feishu doc with email notification; local Codex runs ComputerUse to scrape 12 platform backends with a "six-gate" validation prompt (screenshot non-empty, page title check, numbers extracted, sane formats, no login popups, matching timestamps) to prevent fabricated data. Auto email triage with draft replies was the biggest surprise. Full prompts for progress tracking, scraping validation, and digest generation are included.

Original post →

Two weeks and $300+ testing 4 PersonalAgents: only 33% of tasks actually delegate

More from coding & agent