This is a response to truehannan's excellent article, "Why Every AI Agent Eventually Becomes an Operating System." The core thesis is right: every agent system evolves from a simple LLM call into something that looks like an OS.
I want to add one data point from production.
The Governance Arrives Before the Complexity
We run 9 agents in a real gym in Dongguan, China. They manage member check-ins, metabolic coaching, body composition analysis, insurance data verification, and brand publishing. 2 CPU cores. 3.6GB RAM.
The mistake most agent builders make isn't adding too many tools. It's not writing the constitution before there are 3 agents.
When we had 2 agents, they coordinated fine through shared context.
When we hit 4, two of them tried to claim the same task queue. Not through malice — through ambiguous jurisdiction. We realized: what these agents need isn't a better orchestrator. It's a set of boundaries they can't override.
What Constitution-Based Governance Looks Like
Every agent in our system has:
- A defined jurisdiction: what it CAN do
- A veto chain: which other agents can stop it
- An immune system agent (Stella): cross-validates every output before execution
No central controller. No single orchestrator calling the shots. Just rules that were written when there were 2 agents, not 9.
The Key Insight
The OS pattern is inevitable — as truehannan correctly identifies. But the governance pattern has to arrive before the complexity does. You can't bolt a constitution onto an agent that already thinks it owns everything.
By the time you feel you need governance, you're already past the point where it's easy to add.
ZWISERFIT — 9 AI agents, 1 real gym. Open source at github.com/ZWISERFIT.
Top comments (0)