I have never felt worried about ai.
No matter what happens, life is already beautiful now. If you are in the present then the past or the don’t mater becuse it’s amazing now
In tribute: we celebrate Orange 🍊 and butter 🧈 European or home made by @AnnaLZubarev @iamjuliarosita
1
4
120
Aakashic retweeted
Ok so I now have about 10 personal AI agents working for me, and Grok @bot manages all of them so I can direct them while driving in my @cybertruck 👀
I built the whole setup in one day. Here’s how it works:
- Grok Bot is the main bot. I talk to it and it routes work. Each specialist bot has its own chat, memory and job: coding, research, design, brand, content, my podcast
- My coding bot doesn’t actually write code. It writes a short spec, gives each job its own git branch + worktree, and hands it to Claude Code or Codex
- Claude Code + Codex run headless on Grok Bot’s cloud computer, signed in with my own Claude + ChatGPT subscriptions using the official CLIs. The heavy coding uses those plans instead of my Grok Bot usage
- Claude Code gets UI, copy + refactors. Codex gets backend, tests + bulk edits. Then they review each other’s work
- A separate reviewer bot gates every PR. It checks security, data access, secrets + copy rules, and runs a security scan every week
- An AGENTS.md file acts like the rulebook for both tools. CLAUDE.md imports it
Models switch based on difficulty. Hard or risky work goes to Fable on Claude or GPT-6-Astra on Codex. Normal work goes to Opus or GPT-6.1-Sol, and small jobs go to the cheap tiers. Every Monday a routine updates both CLIs + checks if the labs shipped something better
What I learned:
- Keep it to one bot per workstream.
- The main Grokbot orchestrator writes specs + stays out of the code, so its context stays clean
- I have a different model review the work, and let it actually run the code. Just reading the diff isn’t enough
- Set a rule to isolate one branch per agent. Never let two agents write in the same checkout
- Keep one master rulebook. Multiple instruction files will drift apart
- Close stdin when running Codex headless, or it can sit there waiting for input
- Headless Claude needs permission prompts turned off, so it only runs in throwaway worktrees
- Claude + ChatGPT plan limits are separate pools. When one runs out, the job moves to the matching tier on the other.
- I also had Grokbot build me a live dashboard tracking every plan's usage + annual spend
H/t @gregisenberg for a couple of these ideas from his recent video: a separate reviewer + a twice-daily content scout.
Honestly the wildest part is it all runs while I sleep.
I woke up to a summary of what got shipped overnight 😆
(video generated using @runwayml MCP via @bot)
Art is the answer
d e s I r e d ❤️ is coming…
Art inspired by the symbols of kutch, gujrath India @decorvoyager




















