Capstone: Map Controls to NIST AI RMF and Evaluate Your Agent
A control crosswalk and an honest evaluation, anchored at scale
You'll be able to
- Crosswalk your agent's controls to the NIST AI RMF functions (GOVERN/MAP/MEASURE/MANAGE)
- Evaluate the agent against that crosswalk and name where evidence is missing
Read first
Day 7 Lab
Two honesty notes before you start — say them out loud to your team:
- **The 200 → 500 level ladder is our instructional framing, not an official
Glean certification.** It organizes this curriculum; it confers no credential and is not Glean's.
- **The NIST AI RMF mapping below is instructional synthesis, not a documented
Glean crosswalk.** We map our controls to RMF functions to reason about coverage; Glean does not publish this crosswalk, and you should not present it as Glean's.
Throughout, attribute capability claims to Glean ("Glean describes…"), not to us.
- List every control you built this week: run-mode choice, scoped MCP token, framework tool boundary, layered guardrails, permissions-aware retrieval, audit records.
- Crosswalk each control to one or more RMF functions (use the example below).
- Evaluate: for each function, do you have *evidence* it works, or only intent? Mark the gaps.
- Anchor scale with DBS — Glean reports it freed roughly 10% of work hours across 40,000+ employees — and ask whether your controls hold at that size, not just in a demo.
Working Example: Control → NIST AI RMF Crosswalk
This crosswalk is instructional synthesis, not a Glean-published mapping. Use it to find thin spots — a function with controls but no evidence is your next piece of work. Verified against developer_platform.
CONTROL (built this week) RMF FUNCTION EVIDENCE YOU SHOULD HAVE
──────────────────────────────────────────────────────────────────────────────
Run-mode chosen for the boundary MAP a written trust-boundary note
Scoped MCP token (acts as a user) GOVERN token scope + rotation record
Framework tool = user's access only MAP tool wired to per-user token
Layered guardrails (in/mid/pre-act) MANAGE deny logs showing each fired
Permissions-aware retrieval GOVERN two-user test: different views
Audit record per action MEASURE sampled records, who/what/why
──────────────────────────────────────────────────────────────────────────────
GOVERN: policy & identity in force MAP: context made explicit
MEASURE: behavior is observable MANAGE: unsafe steps are stopped
Scale check (DBS, per Glean): ~10% of work hours freed across 40,000+ people.
At that scale a missing audit trail is not a gap — it is the whole story.The capstone is a MEASURE-and-MANAGE habit: a control with no evidence is a claim, not a control. Find your weakest function, produce one piece of evidence, and you have moved from "we built guardrails" to "we can show they work." That is the difference governance cares about. Verified against developer_platform.
Check for understanding
Which RMF function is your weakest, and what one piece of evidence would strengthen it?
Check yourself
1. Glean describes three agent run-modes. Which list matches them?
2. When a host connects to Glean's MCP server with a scoped token, what happens to access?
3. In a guarded multi-tool agent loop, what is the job of the pre-action gate?
4. Across CrewAI, Google ADK, LangGraph, and the OpenAI Agents SDK, what does every Glean tool have in common?
5. Which statement about this curriculum is accurate?
In 10 seconds
“The capstone: crosswalk your agent's controls to NIST AI RMF, then evaluate honestly where evidence is thin. Two caveats up front — the level ladder and this mapping are our framing, not Glean's official ones.”
Back to all lessons