Enterprise AI Agent Governance & Security Plane
A reference architecture for the three unsolved problems of agentic deployment: identity (zero standing privilege), security (runtime guardrails and a quarantine path), and observability (trace-first debugging). Click any node below to see what it actually covers, and rate your own environment against it — the diagram becomes your maturity map as you go.
The core flow — what happens on every agent action
- •Agent Registry — centralized inventory of every agent, version, owner, purpose, lifecycle state
- •Zero Standing Privilege (ZSP) — agents have zero default permissions; request scoped, ephemeral, task-specific tokens that auto-expire
- •Delegation Mapping — tracks the User → Agent → Agent → Tool chain of custody, proving who authorized what
- •Credential Vault — stores API keys and secrets; agents never hold actual credentials, only short-lived proxy tokens
Agents never receive long-lived credentials.
rejected / ambiguous actions branch here — approved ones loop back to the Policy Decision Point
Cross-cutting layers
Frequently asked questions
What is the Enterprise AI Agent Governance & Security Plane?
A reference architecture for deploying AI agents safely at scale, built around three problems: identity (zero standing privilege — agents hold no long-lived credentials), security (runtime guardrails that enforce every action plus a quarantine path for ambiguous ones), and observability (trace-first debugging, so every action is auditable end to end).
Is this tool free?
Yes, entirely free with no signup. It runs client-side in your browser — nothing you enter is sent anywhere, and your self-assessment scores exist only in that browser tab unless you export them.
What does "zero standing privilege" mean for an AI agent?
It means an agent never holds a long-lived credential at all. It requests a scoped, task-specific, short-lived token from a credential vault when it needs to act, and that token expires automatically — so a compromised or misbehaving agent has nothing standing to abuse.
What is a Policy Enforcement Point (PEP) in this architecture?
The gatekeeper layer that sits between an agent and any real resource. It validates the token, enforces whatever the Policy Decision Point decided, sanitizes sensitive data in the request, and either allows the action (with constraints), denies it with a reason, or routes it to human review — no action reaches a protected resource without passing through one.
Why does quarantine matter as its own component, not just a deny?
A flat allow/deny loses the ambiguous middle case — actions that are not clearly safe but not clearly malicious either. Quarantine stores those, notifies the security team, and lets a human re-enter the action into the policy flow if it turns out to be legitimate, instead of forcing every uncertain case into either a false block or a false allow.