ANTHROPIC NOW SAYS ONE MAIN AGENT WITH SKILLS BEATS A TEAM OF SUBAGENTS. I HAD 6 GROK BOTS PASSING WORK TO EACH OTHER. I MERGED THEM INTO ONE AND IT GOT FASTER AND CHEAPER OVERNIGHT
every handoff between bots lost something. the inbox bot didn't know what the calendar bot promised
what i changed:
1. one main bot plans the day and does the work itself
2. the old bots became skills it loads only when needed: inbox, calendar, research, invoices, posts
3. it keeps one memory file, so nothing gets lost between tasks
4. side bots only for jobs that run in parallel, like watching 3 inboxes at once
the difference in a week:
โ tasks done 31% faster
โ usage down about a third
โ zero "i didn't know you already replied" moments
paste this:
"Be my one main assistant. Plan each task and do it yourself. Turn my old bots into skills you load only when you need them. Keep one shared memory of everything you did. Only start a helper bot for work that has to run in parallel."
full setup in the article โ
I put together one guide for anyone who wants a workflow like the SpaceXSI team's.
One message builds your whole Grok Bot setup. Plus the team's rules, prompt habits, routines and permissions. Every setup is a block you paste.
16-min read. Save it โ https://t.co/gJ8ddzRkkM
JENSEN HUANG'S LITTLE ROBOTS FELL THOUSANDS OF TIMES IN SIMULATION BEFORE THEY EVER WALKED ON STAGE. I STOLE NVIDIA'S TRAINING LOOP FOR MY GROK BOT AND IT HASN'T MESSED UP A REAL TASK SINCE
nvidia doesn't teach robots step by step. they let them fail in a simulator, score every attempt, keep what worked, and only then put them on a real floor.
the same loop, for grok bot:
1. sandbox: it runs a new task on fake copies first. fake inbox, fake calendar, fake clients
2. score: every run gets checked against what i'd call a good result
3. keep the winner: the best version becomes the bot's routine
4. go live: only the tested routine touches my real accounts
5. fall, get up: any real mistake goes back into the sandbox as a new test
first task i trained this way: replying to client emails. 40 rehearsals, then live. zero replies i had to rewrite this week.
paste this:
"Before you do [TASK] for real, rehearse it 40 times on made-up examples. Score each attempt against my rules, keep the best version as your routine, and show it to me. Any mistake you make later becomes a new practice case."
the full training setup is in the article โ
I put together one guide for anyone who wants a workflow like the SpaceXSI team's.
One message builds your whole Grok Bot setup. Plus the team's rules, prompt habits, routines and permissions. Every setup is a block you paste.
16-min read. Save it โ https://t.co/gJ8ddzRkkM