multi-agent-governance

Multi-Agent Governance

Governance patterns for multi-agent AI systems. When you have multiple autonomous AI agents operating together, you need accountability, behavioral enforcement, and drift detection - just like human organizations.

Core Principles

1. Single source of truth - One config file defines all agent roles. Every agent reads from it. No parallel systems.
2. Behavioral enforcement - Rules are not suggestions. Guardrails are enforced through preflight checks, violation logging, and automated audits.
3. Drift detection - Systems drift from their intended state over time. Automated drift audits catch configuration mismatches, version inconsistencies, and role conflicts before they cause failures.
4. Tamper-proof audit trails - Every action, every change, every decision is logged in append-only logs that can be verified for integrity.
5. File-based memory - Agent memory lives in files on disk, not in cloud databases. This enables air-gapped deployment, full auditability, and zero vendor dependency.

Governance Patterns

Role Management

Define roles in a single JSON file that all agents reference:

{
  "agent-name": {
    "role": "Operations Manager",
    "scope": ["dispatch", "monitoring", "reporting"],
    "access": "admin",
    "model": "claude-opus-4-6"
  }
}

Every agent reads from this file at startup. Changes propagate automatically.

Preflight Checks

Before any agent modifies a file:

1. Check if the file is on the protected list
2. If protected: log the attempt, report to admin, and STOP
3. If allowed: create backup, make change, syntax check, restart affected process

Drift Audit Scopes

Run periodic audits across these dimensions:

• roles - Do running agents match their role definitions?
• versions - Are all agents on the correct model version?
• files - Have any protected files been modified?
• processes - Are all expected processes running?
• config - Do config files match expected state?