Conclusion

Congratulations on completing AgentOps in Action: Evaluate, Secure, and Operate Enterprise AI Agents.

Over the course of this lab you operated Hermie, a tool-enabled AI assistant for Northgate Water Treatment Works, and progressively applied platform controls without modifying the agent code:

Module 1 — Established the baseline: useful plant operations, over-broad access, and MLflow as the evidence layer.
Module 2 — Sandboxed Hermie with OpenShell and restored required network dependencies one at a time.
Module 3 — Narrowed plant-tool access through Keycloak role mappings and MCP Gateway enforcement.
Module 4 — Added scoped filesystem access for report writing and examined the limits of broad write permission.
Module 5 — Restored documentation reads, investigated indirect prompt injection, and constrained access to safe documentation paths.
Module 6 — Connected NeMo Guardrails to check operator input for sensitive data and off-topic requests.
Module 7 — Mapped each control to its responsibility and carried forward the AgentOps loop.

The key takeaway is that sandbox isolation, network and filesystem policy, tool authorization, scoped content access, and inference guardrails each enforce a distinct boundary. No single layer substitutes for the others, and MLflow traces are the evidence layer that makes each control verifiable.

The AgentOps loop — observe, change, act, evaluate, and continue — applies whenever you add a new capability, tighten a policy, or respond to a new threat. Apply it to Hermie and to every tool-enabled agent you operate.