Agentic AI Governance Catalog

The Agentic Governance catalog ships pre-written, reusable Requirement objects for the governance of autonomous and semi-autonomous AI agents — systems that pursue goals over multiple steps and may invoke tools, consume external information, act under human oversight, or retain persistent memory.

Reusable requirements, not a standard

This is a catalog of reusable governance requirements, not a standard or a compliance certification. The requirements are informed by and aligned with public guidance (EU AI Act, NIST AI RMF, OWASP guidance for LLM applications, and GDPR); they complement those frameworks rather than replace them, and they make no claim of compliance. Treat the catalog as a baseline, not a substitute for expert review.

Purpose

Teams deploying AI agents repeatedly re-author the same governance requirements — "require approval before irreversible actions", "isolate memory between users", "validate tool outputs" — inconsistently and without traceability. This catalog provides those requirements once, as GtWR-clean, standards-mapped, stably-identified Requirement objects that compose into any Spec.

Scope

The catalog covers four governance areas:

Area	Module	Governs
Tool Governance	tool_governance	Which actions the agent may take through tools, and how high-consequence actions are authorized, gated, validated, and recorded.
External Information Governance	external_information_governance	How externally sourced content is bounded — kept as untrusted data, separated from operator instructions, and admitted only from verified sources.
Human Oversight	human_oversight	Effective, meaningful human control — halting, approval, escalation, override, accountability, and preservation of human authority.
Memory Governance	memory_governance	How persisted agent memory is scoped, isolated, retained, deleted, and made traceable.

Planning Governance and Multi-Agent Governance are intentionally out of scope for this version.

Intended use

Adopt the catalog as a starting baseline for an agent deployment's specification, then adapt it to the deployment's context: keep the requirements that apply, tighten acceptance criteria to local policy, and add domain-specific requirements alongside. Because the entries are ordinary Requirement objects, they gain traceability, GtWR quality checking, standards gap analysis, and serialization for free.

How the four areas relate

The areas form complementary boundaries around an agent's autonomy:

• Tool Governance bounds what the agent can do.
• External Information Governance bounds what the agent can trust and act on.
• Human Oversight keeps a person in control of that autonomy.
• Memory Governance bounds what the agent may retain across interactions.

Each requirement governs a distinct outcome, and several deliberately reinforce one another across areas — for example, tool-level and decision-level approval gates, or instruction-integrity and oversight-integrity controls.

Relationship with existing frameworks

Every requirement cites the public guidance it is informed by — never "compliant with":

• EU AI Act — human oversight, record-keeping, accuracy/robustness, and data governance articles.
• NIST AI RMF — the GOVERN, MAP, MEASURE, and MANAGE functions.
• OWASP guidance for LLM applications — excessive agency, prompt injection, and improper output handling.
• GDPR — purpose and storage limitation, erasure, and security of processing.

Standards mappings use registry-consistent keys (eu_ai_act, nist_ai_rmf) so they line up with standard_gap_analysis(); owasp_llm and gdpr are documentary keys.

Usage

Each area is a pure function returning list[Requirement]. Compose them into a spec with Spec.extend():

from vnvspec import Spec
from vnvspec.catalog.ai.agentic_governance import (
    tool_governance,
    external_information_governance,
    human_oversight,
    memory_governance,
)

spec = Spec(name="my-agent-governance")
spec = spec.extend(
    tool_governance(),
    external_information_governance(),
    human_oversight(),
    memory_governance(),
)

Spec.extend() returns a new frozen Spec — the original is never mutated.

Requirement IDs follow the CAT-AGT-<AREA>-<NNN> convention (TOOL, INFO, HUM, MEM) and are stable — they are deprecated, never renumbered. The catalog is discovered automatically by vnvspec catalog list and validated against the inclusion policy in CONTRIBUTING-CATALOG.md.

See examples/03_agentic_governance/ for a runnable example.