WHAT NEW IN X11.09 MON

AI collections from X · 25 collections · 74,882 posts

personal-agentVIDEO

Building a six-agent farm on an iPhone with open-source assistants

The author runs six open-source agents for inbox triage, morning briefs and end-to-end tasks, with rules that reading runs freely, outgoing actions need approval, and payments or deletes never happen.

Superior
@andreysuperiorArchived 14 days agoOriginally posted 2026.10.26
Superior @andreysuperiorOriginally posted 2026.10.26
Open original on X

THIS IS F*CKING INSANE I turned my iPhone Duo into a secret agent farm. Thorsten's rule, "don't wait for your agents, let them wait on you", is exactly how it runs: 6 open-source agents work at the same time, keep going when I fold the phone at night, and wait for my yes before anything leaves. here's the split: → Tacitus triages the inbox and drafts replies in my voice → Tandem writes the morning brief → Covo runs tasks end to end, self-hosted → Open Duet runs named AI coworkers, each with its own memory (6,100+ stars) → claude-chief-of-staff is my personal AI OS, built on Claude Code (450+ stars) → owl sends Claude Code tool calls to my Telegram for a yes the rules: → reading and sorting run on their own → anything outgoing, a reply or a reminder, waits for my tap → payments and deletes never happen His other point, schedules as the killer feature, is the night shift: each agent runs with claude -p on cron while the phone is folded. Claude Duet is still behind a waitlist. You don't need it for this. start with Open Duet. It needs Node 22+, uv and an engine like Claude Code, Codex CLI or Ollama. No model included. git clone https://github.com/Anil-matcha/open-duet.git cd open-duet/desktop npm run app set approvals to "Ask first", then paste this into Claude Code ↓ "Build my agent farm: 1. One folder per agent with a CLAUDE.md for its job, memory.md and log.md. Each agent reads memory.md first and writes to log.md when done. 2. Root CLAUDE.md rules: reading and sorting run alone, anything outgoing waits for my yes, no payments, no deletes. 3. Add a PreToolUse hook that blocks delete commands, so the rule holds even when an agent forgets it. 4. Add a /morning-brief command in .claude/commands that reads every log.md and shows one screen: done, waiting for me, failed. 5. Add http://night.sh that runs each agent with claude -p, scheduled with cron. 6. Set up owl (pkronstrom/owl-afk) so every outgoing tool call waits for my yes in Telegram. Show me every change as a diff first. No edits until I say go." all 7 repos in the reply ↓ What's the one thing you'd never let an agent do without asking you first? save it for the night you set this up.

This post quotes another post: x.com/thorstenball/status/2114386050638336082

x.com/andreysuperior/status/2114492360025428241Archived 2026.10.26 10:48