Because they show service health, not control effectiveness. An AI platform can be fast and available while agents still use the wrong tools, bypass intended routing, or drive unexpected costs. Governance needs telemetry that links requests to actors, tasks, scopes, and consumption so the organisation can see what the system is doing, not just whether it is responding.
When availability metrics stay green but governance goes blind
Uptime and latency are service-quality signals, which makes them useful but incomplete for agentic ai programmes. They tell you whether the platform is responsive, not whether the agent is acting within approved boundaries, choosing the right tools, or consuming resources in a controlled way. Governance fails when those operational metrics are mistaken for proof of safe autonomous behaviour.
That gap matters because agentic systems can be highly available while still taking harmful or wasteful actions. A low-latency platform can route requests incorrectly, repeat expensive tool calls, or execute tasks outside the intended scope without showing any degradation in traditional infrastructure health.
What those metrics miss in practice
The missing layer is control telemetry. For agentic AI, the key question is not only “did the service respond?” but “who or what initiated the action, under what task context, with which permissions, and against which downstream resource.” Without that linkage, teams cannot tell whether the programme is operating as designed or merely operating quickly.
That is why request-level observability needs to capture actor identity, task intent, scope, tool selection, policy decisions, and consumption patterns. The AI Agent Observability, Audit and Incident Response Guide is useful here because it focuses on attribution and signals that show when an agent has gone wrong, not just whether the platform stayed online. For programmes that need stronger architectural context, Zero Trust for AI Agents shows why per-action verification and removal of standing privilege matter more than availability alone.
At the control layer, AI Agent Authorisation Guide is directly relevant because it treats least privilege, task-scoped access, and human approval as governance mechanisms, not implementation details. If those controls are absent, uptime can remain excellent while the agent still exceeds intended authority.
How to judge whether governance telemetry is actually working
A programme is better governed when its telemetry can answer four questions: what action was taken, which agent or principal took it, what authorisation justified it, and what cost or side effect followed. If those answers cannot be reconstructed after the fact, the programme may be observable at the platform layer but not governable at the decision layer.
That distinction becomes especially important when teams report only service SLOs. High availability can coexist with misrouting, tool misuse, runaway retries, and poor task containment. Useful governance dashboards therefore need to combine execution metadata, approval outcomes, policy denials, and spend or tool-usage trends, so the organisation can detect behaviour drift early.
For programmes with multiple agents or shared orchestration layers, Multi-Agent and A2A Security Guide helps frame why inter-agent delegation and containment need separate visibility. When a single agent can trigger downstream actions in other agents or systems, traditional uptime metrics hide the real control boundary.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST SP 800-53 Rev 5 and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI03 — Identity & Privilege Abuse | Agent governance fails when agents act outside intended authority. |
| Recommendation — Enforce per-action authorization to stop agents exceeding assigned privilege. | ||
| NIST SP 800-53 Rev 5 | AU-6 — Audit Record Review, Analysis, and Reporting | Telemetry must reveal agent actions, decisions, and anomalies. |
| AC-6 — Least Privilege | Governance depends on limiting agent access to task-necessary scope. | |
| Recommendation — Review agent audit records for misuse, drift, and unauthorized actions. Limit agent permissions to the minimum access needed for each task. | ||
| NIST Zero Trust (SP 800-207) | default — Zero Trust Architecture | Per-request verification and no standing trust fit agent governance. |
| Recommendation — Verify each agent request and remove standing trust from autonomous paths. | ||
| OWASP Non-Human Identity Top 10 | NHI-05 — Overprivileged NHI | Agents can be available yet still hold excessive access. |
| Recommendation — Audit and reduce agent privilege to the minimum required scope. | ||
Practitioner Guidance
What to prioritise: Treat governance telemetry as a first-class control objective. If you can only report latency, you have platform monitoring, not agent governance.
What to verify: Confirm that every meaningful agent action can be tied to an actor, a task, a policy decision, and a consumption record. If any of those elements is missing, the control story is incomplete.
What good looks like: Exceptions, denials, unusual tool sequences, and cost spikes should be visible quickly enough to support intervention before the behaviour becomes normalised.
Practitioner takeaway: The right metric for agentic AI is not whether the system stayed up, but whether its actions stayed attributable, bounded, and policy-governed while it was up.