ENGINEERING FIELD NOTE / agentmap

Explainable risk scoring for AI-agent inventories

Stores declared agent metadata and computes a 0-100 risk score from permissions, data access, and tool descriptions — readiness labels and human-readable factors, not runtime enforcement.

Inventory first, score second

AgentMap stores declared agent metadata and computes a deterministic score from permissions, data access, integration type, and tool descriptions. The score is accompanied by factors so a reviewer can inspect why an agent needs attention.

Do not confuse ownership with permissions

The owner field says who is responsible for an agent. It does not grant the agent administrative access. A regression test keeps ownership metadata from inflating permission risk, while sensitive customer-data access is enough to require review.

A readiness label is not enforcement

The current thresholds label scores below 40 ready, 40–74 needs_review, and 75–100 blocked. These are outputs of the prototype rules. A blocked label does not itself revoke permissions or prevent the external agent from running.