MIN’S WORLDWiki
MIN’S GOVERNANCE REFERENCE

Wiki

A practical encyclopedia of agentic AI governance. Browse a subject, look up a term, or follow a learning path.

115 articles · 10 subjects · Updated . Source-based definitions and editorial working definitions are identified on each article. Examples are illustrative.

FIND A NOTE OR CONCEPT

Search Blog & Wiki

⌘ / Ctrl + K

Type a word or phrase to search this site.

Browse by subject

Foundations15 articles →Identity & access20 articles →Execution controls20 articles →Threats & safeguards12 articles →Assurance & operations22 articles →Tools & protocols5 articles →Data & memory8 articles →Cost & FinOps5 articles →Agent efficiency4 articles →System design4 articles →

Learning paths

Govern agent costs

Connect spending to useful outcomes, accountable owners and realistic budgets.

  1. Tokenomics for Agentic AI
  2. Agent Unit Economics & Cost per Successful Task
  3. Agent Cost Allocation, Showback & Chargeback
  4. Agent Cost Forecasting & Scenario Planning
  5. Agent Budget & Loop Controls
  6. Agent Cost Anomalies & Runaway Spend

Improve agent efficiency

Compare routing, caching, latency and reasoning effort without losing quality.

  1. Model Routing & Escalation
  2. Prompt & Tool-Result Caching
  3. Context Window Governance
  4. Agent Latency, Throughput & Queue Design
  5. Reasoning Budgets & Efficiency Frontiers

Design an effective agent system

Choose an architecture, define tools, coordinate work and recover safely.

  1. Agent Task Specifications
  2. Designing an Effective Agentic AI System
  3. Designing Reliable Tools for Agents
  4. Single-Agent, Workflow & Multi-Agent Orchestration
  5. Durable Execution, Checkpoints & Recovery
  6. Agent System Cards

Design accountable delegation

Specify goals, assess affected people and provide meaningful correction.

  1. Agent Task Specifications
  2. Stakeholder Impact Assessment
  3. Proportional AI Governance
  4. Acceptable Use for AI Agents
  5. AI Literacy for Agent Operators
  6. Contestability & Redress

Strengthen agent credentials

Federate identities, bind tokens and control temporary privileges.

  1. Workload Identity Federation
  2. SPIFFE IDs & SVIDs
  3. Token Audience & Resource Binding
  4. Sender-Constrained Tokens
  5. Just-in-Time Agent Access
  6. Agent Permission Drift

Recover reliable workflows

Handle timing, competing changes and ambiguous external outcomes.

  1. Task Deadlines & Cancellation
  2. Retry Budgets & Backoff
  3. Agent Circuit Breakers
  4. Concurrency Control for Agents
  5. Transactional Outbox for Agent Actions
  6. Safe Fallbacks & Degraded Operation
  7. Agent Handoff & Task Ownership

Understand evaluation evidence

Inspect datasets, judges, uncertainty and the claims evidence supports.

  1. Synthetic Data for Agent Evaluation
  2. Evaluation Contamination
  3. LLM-as-a-Judge Governance
  4. Uncertainty & Calibration
  5. Assurance Cases for AI Agents
  6. Agent Governance Metrics

Protect the agent supply chain

Review generated dependencies, tool inputs and persistent context.

  1. Software Supply Chain for Agents
  2. Package Hallucination in Coding Agents
  3. Tool Schema Security
  4. SSRF in Agent Fetching Tools
  5. Agent Memory Poisoning
  6. Data Lineage in Agent Workflows
  7. Deletion Across Agent Memory & Derived Data

Start with governance

Purpose, ownership, risk and bounded autonomy.

  1. AI Governance
  2. Agentic AI Governance
  3. Agent Inventory
  4. Agent Risk Assessment
  5. Autonomy Boundaries
  6. Agent Accountability

Understand an access decision

Identity, task permissions, current facts and enforcement.

  1. Authentication
  2. Authorization
  3. Purpose-Bound Authorization
  4. Policy Information Point
  5. Policy Decision Point
  6. Policy Enforcement Point
  7. Agent Credential Lifecycle

Operate and improve

Find failures, retain evidence and respond.

  1. Prompt Injection
  2. Evaluations & Red Teaming
  3. OpenTelemetry
  4. Observability
  5. Audit Trail
  6. Incident Response

Secure memory and retrieval

Persistent context, source access and permitted data flows.

  1. Agent Memory Governance
  2. Retrieval & RAG Security
  3. Prompt Injection
  4. Agent Data Exfiltration
  5. Audit Trail

Control autonomous workflows

Bound delegation, budgets, retries and downstream effects.

  1. Autonomy Boundaries
  2. Delegated Authority
  3. Multi-Agent Security
  4. Agent Budget & Loop Controls
  5. Idempotency & Compensation
  6. Revocation & Task Cancellation

Review and release agents

Supplier evidence, material changes and ongoing evaluation.

  1. Agent Risk Assessment
  2. Agent Vendor Due Diligence
  3. Tool Poisoning
  4. Agent Change Management
  5. Continuous Agent Evaluation

Build enforceable boundaries

Threat modeling, service identities, tenant separation and tested controls.

  1. Agent Threat Modeling
  2. Zero Trust for AI Agents
  3. OAuth Token Exchange for Agents
  4. Tenant Isolation for AI Agents
  5. Agent Output Validation
  6. Policy as Code Testing

Make oversight useful

Review consequential actions, protect evidence and handle exceptions.

  1. Browser Agent Governance
  2. Approval Fatigue & Human Oversight
  3. Telemetry Data Minimization
  4. Audit Evidence Integrity
  5. AI Risk Exceptions

Ship, recover and retire

Evidence for release, controlled expansion and responsible closure.

  1. Agent Evaluation Datasets
  2. Agent Artifact Provenance
  3. Staged Agent Rollout
  4. Emergency Access for Agent Operations
  5. Agent Decommissioning
READ THE ORIGINALS

PDF reading library

Open official AI governance publications in a full-screen reader. With Google sign-in, save page notes, pen marks and highlights to your reading shelf.

AI Risk Management FrameworkNIST · 2023 · Governance foundationsRead and annotate →Generative AI ProfileNIST · 2024 · Generative AI governanceRead and annotate →AI Agent Identity & AuthorizationNIST NCCoE · 2026 concept paper · Agentic AI governanceRead and annotate →
Explore the whole library →

Article directory

115 concepts

Foundations

Acceptable Use for AI Agents

Translate organizational boundaries into understandable rules and enforceable behavior.

Agent Accountability

Connect agent actions to accountable owners, decision rights, evidence, and escalation.

Agent Inventory

An agent inventory is a maintained record of deployed or proposed agents, their owners, purposes, dependencies, permissions and review status.

Agent Risk Assessment

Assess an agent’s actions, affected people, dependencies, and remaining risk before and during operation.

Agent Task Specifications

Define goals, constraints, completion evidence and escalation before execution.

Agentic AI Governance

Governance for AI systems that plan, call tools, and act across workflows.

AI Governance

Ownership, decisions, and oversight throughout an AI system’s lifecycle.

AI Literacy for Agent Operators

Build role-specific knowledge and practiced intervention skills.

AI Risk Exceptions

Make departures from internal controls explicit, bounded, owned and subject to reconsideration.

Autonomy Boundaries

Describe autonomous action through explicit permissions, thresholds, and intervention points.

Contestability & Redress

Make explanations, human review and correction available to affected people.

Proportional AI Governance

Match controls and review effort to consequences, exposure and uncertainty.

Reward Hacking & Proxy Objectives

Examine the gap between a measurable objective and the result people actually want.

Semantic Interoperability

Semantic interoperability means that systems can exchange data and preserve the intended meaning of its concepts, measures and relationships. A shared model format helps exchange definitions, but teams must still agree on what a business term means in each context.

Stakeholder Impact Assessment

Identify affected groups, competing interests and consequences before choosing controls.

Identity & access

Agent Credential Lifecycle

Manage credentials from provisioning through expiry, rotation, compromise response, and retirement.

Agent Identity

A verifiable identity for the software actor, linked to its owner and task.

Agent Permission Drift

Compare intended authority with effective permissions as teams, tools and environments change.

Attribute-Based Access Control

Attribute-based access control evaluates rules using properties of the actor, resource, requested operation and environment.

Authentication

Authentication establishes confidence in the identity claimed by a person, process or device. For an agent, it verifies the actor presenting a credential.

Authorization

Authorization determines whether an identified actor may perform a particular action on a resource under applicable rules.

Capability-Based Access for Agents

Reason about narrowly scoped access handles, forwarding risks and revocation.

Delegated Authority

In agent governance, delegated authority is a bounded grant to act for a requester on a defined task, with specified actions, resources and limits.

Emergency Access for Agent Operations

Define exceptional access with accountable activation, narrow scope, automatic expiry and review.

Just-in-Time Agent Access

Grant temporary authority with explicit scope, activation conditions and expiry.

Least Privilege

Least privilege limits an actor’s access to what its assigned work requires.

OAuth Token Exchange for Agents

Understand subject and actor identities, downstream credentials and the limits of token exchange.

Purpose-Bound Authorization

Limit a grant by task, resource, action, and conditions instead of broad standing access.

Secrets Management for AI Agents

Keep sensitive credentials in controlled execution paths and manage their complete lifecycle.

Sender-Constrained Tokens

Bind token use to key material while recognizing what this does and does not prevent.

Separation of Duties in Agent Workflows

Keep conflicting responsibilities separate across people, agents and service accounts.

SPIFFE IDs & SVIDs

Distinguish a workload identifier from its verifiable identity document and business permissions.

Token Audience & Resource Binding

Constrain where an access token is accepted and verify the intended recipient at the resource.

Workload Identity Federation

Establish trust between identity systems while constraining which workloads can receive access.

Zero Trust for AI Agents

Apply resource-specific access decisions to agents across internal services, tools and delegation paths.

Execution controls

Agent Budget & Loop Controls

Bound spend, tokens, steps, retries, time, and parallel work with enforceable operating limits.

Agent Circuit Breakers

Interrupt repeated failures and reopen access through controlled recovery.

Agent Output Validation

Validate syntax, meaning, authority and rendering context before a model result reaches another system.

Approval Fatigue & Human Oversight

Design agent review requests that expose consequences, preserve control and avoid routine rubber-stamping.

Approval-to-Execution Races

Prevent state changes between review and commitment from invalidating an approval.

Concurrency Control for Agents

Protect shared state from stale writes and conflicting decisions.

Human Approval & Oversight

Human oversight assigns people meaningful opportunities and authority to review, intervene in or stop AI-supported work. Approval is one specific oversight mechanism.

Idempotency & Compensation

Prevent duplicate effects and define recovery for multi-step agent actions that cannot simply be undone.

Policy Administration Point

A policy administration point creates and manages access-control policies used by the decision system.

Policy as Code Testing

Test authorization rules, missing context and real enforcement before policy changes reach agent workflows.

Policy Decision Point

A policy decision point evaluates applicable access-control policies and produces an authorization decision.

Policy Enforcement Point

A policy enforcement point requests an access decision and enforces the result at the resource-access boundary.

Policy Information Point

The PIP supplies attributes to a policy evaluation; it does not make the decision.

Retry Budgets & Backoff

Limit retries across the whole workflow and avoid turning temporary failures into overload.

Revocation & Task Cancellation

Revocation withdraws previously granted access. Task cancellation stops further workflow execution; the two must be coordinated because cancelling a task does not automatically invalidate every credential.

Runtime Governance

Policy checks, enforcement, and evidence while an AI system is operating.

Safe Fallbacks & Degraded Operation

Choose restricted operating modes that preserve permissions and make missing evidence visible.

Task Deadlines & Cancellation

Distinguish waiting limits from cancellation and propagate task deadlines through downstream work.

Tool Boundary

A tool boundary is the point where an agent’s proposed action becomes an operation on an external system. It is a useful place to enforce authorization and validate parameters.

Transactional Outbox for Agent Actions

Coordinate state changes and event publication without assuming exactly-once processing.

Threats & safeguards

Adversarial Robustness for Agent Workflows

State the threat model and test whole-workflow behavior under deliberate manipulation.

Agent Data Exfiltration

Control external destinations and data flows across tools, messages, artifacts, and telemetry.

Agent Memory Poisoning

Protect durable context from malicious or mistaken updates that influence later behavior.

Agent Threat Modeling

Map adversarial paths across prompts, tools, memory, users and downstream services.

Confused Deputy

Prevent a privileged intermediary from exercising its access for an unauthorized caller or purpose.

Excessive Agency

Excessive agency is risk created when an LLM application can take actions through more functionality, permissions or autonomy than its task needs.

Package Hallucination in Coding Agents

Verify generated dependency suggestions before installation or release.

Prompt Injection

Prompt injection is an attempt to make a model treat lower-trust input as instructions that override the task or the application’s rules. It can change an answer, redirect a workflow or induce an unauthorized tool action.

Sandboxing

Sandboxing runs software in an environment that restricts its interactions with the surrounding system.

Software Supply Chain for Agents

Review dependencies, connectors and updates as part of the agent’s security boundary.

SSRF in Agent Fetching Tools

Constrain network destinations independently of the model’s interpretation of a link.

Tool Poisoning

Review tool metadata and returned content as possible instruction-injection surfaces.

Assurance & operations

Agent Artifact Provenance

Connect a deployed agent to its build, dependencies, prompts, policies and release evidence.

Agent Change Management

Review material changes to models, prompts, tools, memory, permissions, and operating boundaries.

Agent Decommissioning

Retire agents by closing execution paths, removing access and resolving retained data and downstream work.

Agent Evaluation Datasets

Document cases, expected outcomes, coverage gaps and dataset changes before trusting an agent score.

Agent Governance Metrics

Measure outcomes and control failures with clear denominators, context and decision owners.

Agent Incident Tabletop Exercises

Rehearse containment, ownership and communication using a realistic unfolding scenario.

Agent System Cards

Document the deployed system’s purpose, authority, evidence and operational limits.

Agent Vendor Due Diligence

Review third-party agents and tools for data handling, access, change control, evidence, and exit options.

Assurance Cases for AI Agents

Connect bounded claims to evidence, assumptions and unresolved objections.

Audit Evidence Integrity

Protect evidence from alteration and distinguish recorded claims from independently observed outcomes.

Audit Trail

An audit trail is a chronological record used to reconstruct relevant actions and decisions. For agents, it should connect the requester, authority, tool operation and outcome.

Continuous Agent Evaluation

Repeat targeted evaluations as agent configurations, inputs, integrations, and operating conditions change.

Evaluation Contamination

Protect evaluation independence and disclose known or uncertain exposure to test material.

Evaluations & Red Teaming

Evaluations measure a system against stated criteria. Red teaming deliberately probes ways it may fail or be misused, complementing ordinary performance checks.

Incident Response

Incident response is the organized handling of security incidents, including detection, response and recovery within ongoing risk management.

LLM-as-a-Judge Governance

Validate automated judges against task rubrics, independent evidence and human review.

Model Cards in Agent Procurement

Read model evidence critically and identify what still needs deployment-specific evaluation.

Observability

Observability uses emitted signals to understand a running system’s behavior. Agent workflows benefit from correlated traces, metrics and logs across tool calls.

OpenTelemetry

OpenTelemetry is an open-source, vendor-neutral framework for instrumenting software and generating, collecting, and exporting telemetry such as traces, metrics, and logs. It supplies APIs, SDKs, conventions, a protocol and an optional Collector; storage, dashboards and alerting come from other systems.

Staged Agent Rollout

Move from controlled evaluation to limited operation using explicit expansion and stop criteria.

Telemetry

Telemetry is data a running system emits about its activity. Traces connect operations in a request, metrics measure behavior over time, and logs record events. These signals help operators investigate agent workflows when they include useful context.

Uncertainty & Calibration

Validate confidence signals before using them to automate decisions or trigger escalation.

Tools & protocols

Agent Handoff & Task Ownership

Transfer task state and responsibility without silently transferring unrestricted authority.

Browser Agent Governance

Control browser sessions, destinations, uploads and consequential clicks while handling untrusted page content.

Model Context Protocol

Model Context Protocol is a protocol for connecting AI applications with external capabilities and context, including tools, resources and prompts.

Multi-Agent Security

Preserve identity, task scope, data boundaries, and accountability across agent-to-agent handoffs.

Tool Schema Security

Keep tool schemas narrow, versioned and consistent with execution-time validation.

Data & memory

Agent Memory Governance

Control the purpose, provenance, permissions, retention, and integrity of persistent agent memory.

Context Window Governance

Manage selection, summarization and truncation of the information an agent sees.

Data Lineage in Agent Workflows

Trace information through retrieval, transformations, summaries and downstream decisions.

Deletion Across Agent Memory & Derived Data

Coordinate deletion and access removal across memory, indexes, caches and retained evidence.

Retrieval & RAG Security

Protect retrieval sources, document permissions, context integrity, and tenant separation.

Synthetic Data for Agent Evaluation

Use synthetic cases deliberately while checking realism, coverage and privacy claims.

Telemetry Data Minimization

Design agent traces around operational questions, controlled content capture and limited access.

Tenant Isolation for AI Agents

Preserve tenant boundaries across tools, retrieval, caches, queues and long-term memory.

Cost & FinOps

Agent Cost Allocation, Showback & Chargeback

Attribute agent spending across parent tasks, delegated work and shared services without double counting.

Agent Cost Anomalies & Runaway Spend

Detect unusual spending, identify its cause and contain it without losing track of in-flight work.

Agent Cost Forecasting & Scenario Planning

Forecast task mix and execution behavior, with explicit assumptions and uncertainty.

Agent Unit Economics & Cost per Successful Task

Connect agent spending to accepted outcomes, quality and the cost of operating the whole workflow.

Tokenomics for Agentic AI

Understand token billing, repeated context, caching and agent loops, then connect consumption to verified outcomes.

Agent efficiency

Agent Latency, Throughput & Queue Design

Measure end-to-end completion, waiting and capacity rather than only model response speed.

Model Routing & Escalation

Route work by measured capability and constraints while accounting for routing and retry overhead.

Prompt & Tool-Result Caching

Reduce repeated work while preserving freshness, access boundaries and correct billing assumptions.

Reasoning Budgets & Efficiency Frontiers

Evaluate marginal improvements and stop unproductive work within explicit resource limits.

System design

Designing an Effective Agentic AI System

Choose an architecture from evidence about the task rather than from the number of agents or tools.

Designing Reliable Tools for Agents

Design clear operations, bounded inputs and informative results that reduce both errors and wasted calls.

Durable Execution, Checkpoints & Recovery

Persist meaningful state and recover deliberately across model calls, approvals and tool effects.

Single-Agent, Workflow & Multi-Agent Orchestration

Choose sequential, parallel or delegated execution according to dependencies and verifiable outputs.

A–Z term index

Expand +Collapse −

A

B

C

D

E

H

I

J

L

M

O

P

R

S

T

U

W

Z