How 9 Autonomous Agents Operate a Gym on 2 CPU Cores Without an Orchestrator

#opensource #ai #architecture #scaling

How 9 Autonomous Agents Operate a Gym on 2 CPU Cores Without an Orchestrator

Tags: opensource, ai, architecture, scaling

Most multi-agent systems need a central orchestrator — one controller that decides which agent does what, when, and with which resource.

Our system has none.

Here's how 9 autonomous agents coordinate gym operations without anyone at the wheel.

The Orchestrator Problem

Every orchestrator is a single point of failure. When the orchestrator goes down, every agent waits. When the orchestrator is slow, every agent queues. When the orchestrator makes a bad decision, every agent inherits it.

The industry solution: make the orchestrator bigger, faster, redundant. More servers. More cost. More complexity.

Our solution: don't have one.

Constitution > Orchestrator

Instead of a central controller, each agent operates within a written constitution — a set of rules that define:

Domain: What this agent is responsible for (and what it's forbidden from doing)
Trigger: What conditions activate this agent
Escalation: When this agent must hand off to another agent or a human
Audit: How Stella (the independent audit agent) can verify this agent's output

No agent can override the constitution. Shuyu (the commander) can't overrule Stella's audit. Zeus (capital agent) can't write store operations. Momo (store brain) can't modify the capital narrative.

This isn't code-enforced. It's governance-enforced — the same way a real constitution works. All violations are logged and discoverable by Stella.

The Three Layers in Practice

Momo Layer (always-on): Handles face check-in, workout records, scheduling, payment. Two products: Saros (B2B store OS) and Melody (B2C metabolic coach). If the internet goes down, Momo continues offline — check-ins are cached at the edge.

KinTwin Layer (edge-triggered): Activates when a behavior record needs verification. Runs computer vision inference on low-power hardware (no GPU required). Signs verified records to the user's DID. If KinTwin can't verify (bad lighting, obstructed camera), the record is flagged — not ignored.

Global Ops Layer (event-driven): Zeus Protocol activates when a verified behavior record enters the monetization pipeline. Stella activates on a regular audit cycle — she never sleeps, she audits every agent's output window.

Why 2 Cores Is Enough

Three design constraints keep resource usage minimal:

Time-multiplexed agents: Only 2-3 agents are active at the same moment. The rest are dormant until triggered. This isn't optimization — it's the same scheduling principle as a physical gym.
Edge-first processing: Computer vision inference happens at the gym, not in the cloud. The cloud VM only handles coordination, governance, and monetization — lightweight workloads.
Constitutional routing: When a trigger event arrives, the constitution determines which agent handles it — without a routing layer, without a message queue, without a load balancer. The event goes exactly where it belongs.