← Systems LibraryEditorial lens: Systems Architect

How to Run an AI Swarm Without Losing Control

A hundred agents do not create leverage by themselves. They create a coordination problem. The operating layer is where the real system begins.

Start with three layers

1. Brain / Policy
Goals, source of truth, permissions, definitions of done, evidence requirements, cost rules and human gates. This layer should survive model changes.
2. Runtime / Harness
The agent products or model runtimes actually executing tasks. These can change frequently. Treat them as replaceable workers, not the company memory.
3. Execution
Browsers, repositories, databases, payment systems, publishing systems, inboxes and production environments where actions have real consequences.

Six controls that matter more than agent count

Role ownership. Every agent should own a named job. Overlapping ownership creates silent conflict and duplicate work.
Shared state. Decisions, current status and handoff context must live outside any one conversation or model.
Permission boundaries. Separate observe, draft, recommend, execute and irreversible-change permissions.
Human gates. Money movement, publishing, deletion, production changes and other irreversible actions should have explicit approval rules.
Retry and circuit-break rules. Repeated failure should stop, not turn into an infinite loop that spends tokens or damages an account.
Evidence. Agents should leave source links, timestamps, action logs and anything they could not verify.

Why this matters now

Recent AI-agent products increasingly support persistent computers, routines, shared files and parallel Bots. That lowers the cost of building desk-like structures around one operator. The failure mode is predictable: organizations add more agents before defining policy, state and escalation.

The result is “split brain”: two capable agents follow different assumptions, use different sources of truth, or execute incompatible actions.

A minimal operating contract

GOAL — What outcome are we optimizing?
SOURCE OF TRUTH — Which system wins when data conflicts?
CAN DO — What may the agent execute alone?
MUST ASK — Which actions require approval?
STOP — What conditions halt the workflow?
EVIDENCE — What must be recorded before work counts as complete?
STATE — What must be handed to the next agent or runtime?

The portable part is the asset

Models will keep changing. The durable asset is the operating contract around them: Brain, Policy, Skills and State. If that layer is portable, a company can replace runtimes without rebuilding its operating memory from scratch.

Cross-Agent Operating Kit →

The product link is optional reading for teams that want reusable implementation templates. This article stands on its own. Named AI products are referenced only for compatibility and industry context; Stratum Praxis is independent of those vendors.