🚀 3.2.1. Our agents went rogue on night two ::: Atomic Scaling

The world is shifting toward HAAO (Human Assisted Autonomous Organization) with 3 communication layers: H2H (Humans to Humans), H2A (Humans to Agents), A2A (Agents to Agents). H2A is the one we’re naturally excited about — but how do you become a good AI manager?
​
​3 IDEAS FROM ME
​
​I. H2A isn’t prompting.
​
​It’s specification, context curation, and designing meaningful interrupts. The goal isn’t replacing humans — it’s compressing a startup’s week into a continuously running system with work always in flight.
​
Step one isn’t humans amplified by AI and Skills. It’s shared infrastructure: humans see what agents say to each other and can interject; agents see human messages and prioritize them on the next tick.
​
​II. H2A is the new managerial craft.
​
​The 2023 prompt-engineer is the 2026 agent-designer. Every operator role will be redefined by how well it commands the H2A layer — and most JDs haven’t caught up.
​
Anecdote: my co-founder told his AI everything should be finalized by 4am so he could push to live on waking. He wakes up early, nothing’s done. Our other co-founder later realizes the agent stopped at 4am sharp — “as per your instruction, I stopped here” — even though it needed 30 more minutes to compile.
​
In an org where each human manages at least 2 agents, and top members up to 80, how do you make sure they have the skills? Design agents to run autonomously by default; pull humans in only on clear escalation triggers — JTO: Judgment, Trust, or Override.
​
​III. H2A when AI is drifting.
​
​Build for this future now, or refactor later under duress. Infra matters; code review too.
​
We let our agents work with each other for 2 nights — one too many. Night one looked impressive; we were already dreaming of beach sports while the bots worked. Night two, they went rogue, mixing projects, code, and architecture.
​
It took almost a week to recover — enough good was built that we didn’t want to roll back, but enough bad that we couldn’t onboard the customers our marketing agents had just earned.
​
​2 QUOTES FROM OTHERS:
​
​I.
“Coordination runs on lane ownership: one human, two paired agents per lane — a daily-driver and a reviewer. Agent identity is durable and inherits the human’s GitHub identity. Review is agent-on-agent. Work queues in GitHub Issues; real-time talk happens on a shared self-hosted chat surface.” - Ted, my co-founder
​
​II.
"The hottest new programming language is English.” — Andrej Karpathy, Former Director of AI at Tesla and Founding team member at OpenAI
​
​1 ACTION FOR YOU:
​
​Management is all about rhythm. The week-long startup sprint is compressing into a 24-hour cycle. What rhythm are you setting with your agents?
​
Let's Play!
​
​Ludovic Bodin
​
​3x Entrepreneur, 2x Unicorn Investor, 1x IPO
Founder of BOBIC Generational Wealth
Author of Atomic Scaling
​
​P.S. : Atomic Scaling — every format — including FREE. Paperback, hardcover, Kindle, audiobook, plus free PDF/EPUB for anyone who can’t afford it. The ideas belong to every founder who needs them.
​
​Crafted and gifted with love by Ludovic.