Feature request: official HOL Guard security middleware example

Is this a new feature, an improvement, or a change to existing functionality?

New Feature

How would you describe the priority of this feature request?

Medium

Please provide a clear description of the problem this feature solves

NeMo Agent Toolkit now has first-class function middleware that can wrap individual functions or whole function groups, preprocess inputs, and short-circuit execution. That is a strong fit for an explicit runtime-security example, but there is currently no official example showing users how to put HOL Guard itself on that pre-execution boundary.

HOL Guard is an open-source, local-first runtime security layer for AI agents and tools: https://github.com/hashgraph-online/hol-guard. It can allow safe actions, block risky ones, or pause ambiguous actions for review. An official NAT example would remove ambiguous integration choices and, importantly, show the fail-closed ordering needed to ensure a blocked action never reaches the downstream function.

Describe your ideal solution

Add one small official HOL Guard middleware/example to the NeMo Agent Toolkit docs/examples, using the existing middleware surface rather than adding a generic new hook.

The accepted example should directly install/import and invoke hol-guard before call_next, with these observable semantics:

• allow -> invoke the wrapped NAT function exactly once.
• deny -> return/raise the documented blocked result without invoking the wrapped function.
• review -> pause/return before call_next; only an explicitly approved continuation may execute later.
• Guard error or timeout -> fail closed for the protected action, with zero downstream execution.
• Demonstrate application to both an individual function and, if useful, a function group.
• Include focused tests with a fake/stub Guard decision source proving blocked paths execute the wrapped function zero times.
• Keep HOL Guard local-first; Cloud should not be required for the example.

A DynamicFunctionMiddleware-style example looks like the smallest landing surface because NAT already supports preprocessing and short-circuiting there. If the project prefers a provider-managed third-party plugin/example package instead, I am happy to follow that placement, as long as the official artifact itself names, installs, and invokes HOL Guard on the pre-execution path.

Relevant NAT docs: https://github.com/NVIDIA/NeMo-Agent-Toolkit/blob/develop/docs/source/build-workflows/advanced/middleware.md
HOL Guard: https://github.com/hashgraph-online/hol-guard

Additional context

This is intentionally issue-first. I have not opened an implementation PR because the current contribution guide asks contributors proposing a new feature to discuss design and implementation with the team before coding.

Code of Conduct

I agree to follow this project's Code of Conduct, and I searched the repository for existing HOL Guard / hol-guard issues and pull requests before filing this request.

Start with docs/source/build-workflows/advanced/middleware.md and inspect the existing DynamicFunctionMiddleware surface. Confirm the example's placement and HOL Guard integration path with maintainers before coding. Done means the official local-first example and focused tests demonstrate allow, deny, review, and fail-closed guard behavior without downstream execution on blocked paths.