Day 5: The Hypervisor

Overview

Days 1-4 taught SAAC — Single-Agent Agentic Coding: one human, one agent, one session, done well. Day 5 is the step up to MAAC — Multi-Agent Agentic Coding: one human hypervisor running a fleet of agents at once, each an independent node coordinated through a shared whiteboard. SAAC measures how autonomous one agent is; MAAC measures how many agents one human can run. They are different axes, and the second is built on top of the first.

That second axis hides a maturation. Early on, agentic coding feels like a prompting skill: find the right words, get the right output. With experience it inverts — the rails and rigging you build around the words come to matter more than the words themselves. Telling an agent what to do is the easy part — the vibe-coding end of the spectrum — and at fleet scale it is a liability: a literal-minded junior who is only told what to do will hand you, confidently, something that runs and is wrong. Work that runs is one thing; work that is correct is another, and harder. Correctness does not come from the instruction; it comes from the rigging you assemble around it: the tests, the types, the steel thread, the independent verifier, the single-writer board — the guardrails the machine itself can run against to confirm that what it built both works and is right. Most of Day 5 is the building of that rigging — the hypervisor's real product is not prompts but the scaffolding that makes a fleet's output trustworthy.