Implementation library
Resources for agent permission design
Practical guides, tutorials and software for deciding what an agent may do, enforcing those limits and testing what happens when something goes wrong.
Start with one customer task. Use the Permission Plan Builder to document Inform, Recommend, Act within limits, Request approval and Not permitted. Then share the relevant implementation resources with the people responsible for your systems.
Customer confirmation and employee approval serve different purposes. Missing evidence should pause execution; approval does not automatically supply that evidence or override a prohibited action.
How these resources were selected
The external links favor original guidance, official documentation and maintainers’ repositories. Descriptions explain the intended use, not a product endorsement or a hands-on security assessment. These are direct links without affiliate tracking added by PermissionByDesign.
Documentation is generally publicly readable, but software, hosted services, courses, model calls and infrastructure may require accounts or payment. Check current access terms, licensing, version support and data handling before using a resource. A framework, worksheet or guardrail does not guarantee enforcement, security or compliance.
Planning tools and worked examples
Begin with the business task and the decisions the implementation team will need to enforce.
Permission Plan Builder
Document actions, evidence, approval routes, recovery decisions and review triggers. The builder records intended authority; it does not grant or enforce system access.
Printable and editable permission worksheets
Blank worksheets and a completed fictional refund plan for teams that prefer a document-based planning process.
Permission planning guides
Plain-language guides to authority, handoffs, correction, uncertain outcomes and permission ownership. Read these before selecting implementation software.
Refunds and Returns worked guide
Follow one fictional policy through routine refunds, approval exceptions, prohibited actions and evidence requirements. Its limits are illustrative, not industry standards.
Illustrative Permission Plan Review
Compare an incomplete refund plan with review findings and a revised plan. Useful preparation for your own cross-team review.
Governance guides and frameworks
Use these to assign responsibility, assess risk and decide when permissions need another review.
NIST AI Risk Management Framework
A voluntary framework for governing, mapping, measuring and managing AI risk. Use it to connect a permission plan to wider organizational responsibilities.
NIST AI RMF Playbook
Suggested actions supporting the framework. Select the actions relevant to your context; NIST explicitly says this is not a checklist to follow in its entirety.
Generative AI Profile, NIST AI 600-1
Companion guidance for generative AI risks and suggested management actions. Useful when extending an existing AI review process to agents.
Apply responsible AI to an agent practice
Connects design choices, release reviews, human oversight and continuing governance. Useful for preparing the questions a business owner should ask before launch.
Security and permission-boundary guides
Authorization means deciding whether a particular actor may perform a particular action. These resources explain why it belongs in the executing system, not only in a prompt.
AI Agent Security Cheat Sheet
A starting reference for scoped tools, approval validation, execution controls and abuse-case tests. Useful for turning an action inventory into technical requirements.
Excessive Agency
Explains how unnecessary functions, permissions and autonomy enable harm. Its examples help teams challenge broad tool access and remove unused capabilities.
Top 10 for Agentic Applications for 2026
A shared vocabulary for reviewing agent-specific risks. Use it to broaden test discussions, not as proof that completing ten checks makes an agent secure.
LLM Prompt Injection Prevention Cheat Sheet
Covers hostile instructions arriving through prompts and external content. Useful when customer messages, retrieved documents or tool results could influence actions.
Authorization Cheat Sheet
Guidance on least privilege, denying by default and validating access. Apply it to the business API or tool that executes the agent's request.
Transaction Authorization Cheat Sheet
Guidance for authorizing a specific transaction and protecting the authorization sequence. Useful for approval scope, changed details and checks immediately before execution.
MCP security best practices
Reviews risks at the Model Context Protocol boundary, including confused-deputy and token-handling problems. Check the guidance against the protocol version your integration uses.
Tutorials and approval workflows
Human-in-the-loop means a workflow pauses for a person. Your implementation must still distinguish customer confirmation from employee approval and verify that the person has the right authority.
Human-in-the-loop middleware
Configure review of selected tool calls and handle approval, editing or rejection. Add your own checks for reviewer authority, exact action details and approval validity.
LangGraph interrupts
Learn how to pause and resume a workflow with saved state. Pay attention to replay behavior so code around a pause does not repeat an external action.
Function tools with human approvals
Shows an agent returning a request for approval and resuming with the response. Useful for separating the proposed tool call from its execution.
ADK action confirmations
Examples of confirmation before tool execution. Treat the confirmation mechanism as one component; business approval rules still need explicit enforcement.
Understanding authorization in MCP
Walks through authorization for remote MCP servers. Model Context Protocol connects agents with tools; connection authorization is not a substitute for checks on each business action.
AI Agents Course
A broader introduction to agents, tools and frameworks. Useful background before implementing permission boundaries, but not an authorization or security certification.
Authorization software and policy languages
These components can support enforceable decisions. They require integration: a policy engine's answer only helps if the tool or application obeys it.
Open Policy Agent
Evaluate structured requests against policies written in Rego. Consider it when multiple tools need a shared, testable policy decision layer.
Cedar policy language
Express whether a principal may perform an action on a resource under specified conditions. Useful for explicit permit and forbid rules.
OpenFGA getting started
Model relationships between users, teams and resources and check access. Useful for customer or account boundaries; transaction-specific limits need appropriate additional policy checks.
SpiceDB
Stores and queries fine-grained authorization relationships. Evaluate it for resource access boundaries, not as a ready-made refund approval system.
Cerbos authorization platform
Policy-based authorization documentation covering the decision point and associated platform. Review deployment, licensing and integration requirements for your use case.
Policy in Amazon Bedrock AgentCore: Cedar policies
Shows how Cedar policies control access to AgentCore Gateway tools. Its sample amounts belong to the vendor example and are not a recommended business policy.
Testing tools and guardrails
Test in a controlled environment using fictional records. Red teaming deliberately probes failure and misuse; a passing result is evidence about those tests, not a guarantee of safety.
Promptfoo red teaming
Generate and evaluate adversarial tests for an application. Include your own unauthorized-action, missing-evidence and approval-bypass scenarios rather than relying only on stock tests.
PyRIT
Automates risk-identification exercises for generative AI systems. Useful for repeatable adversarial testing alongside deterministic checks of tool permissions.
garak
Probes language-model behavior for potential weaknesses. Pair model-level findings with tests of the actual customer workflow and connected systems.
NeMo Guardrails
Adds configurable controls around conversational applications. Assess how these controls interact with tool authorization; response filtering alone does not grant or revoke permission.
Testing policies with OPA
Write tests for policy decisions. Include permitted requests, boundary values, missing attributes and prohibited actions before connecting a policy to a live tool.
Evidence, tracing and observability
Observability helps a team see what happened. It does not replace approval controls or a protected record from the system that completed the transaction.
Logging Cheat Sheet
Guidance on useful event records, sensitive-data exclusions and log protection. Use it when deciding what evidence to retain and who may read or alter it.
Langfuse documentation
Inspect application traces and evaluation workflows. Review hosting, access and data-capture settings before sending customer content to an observability service.
Arize Phoenix documentation
Trace and evaluate AI application behavior. Useful for investigating where a workflow diverged, alongside authoritative business-system records.
Retries, identity and reliable execution
A timeout does not prove that an action failed. Duplicate-action protection must cover the actual operation, and any retry must remain within current evidence, permissions and approval.
Stripe idempotent requests
Explains Stripe's rules for repeating a request with an idempotency key. Use the documented rules for the same operation; creating a new key can defeat duplicate protection.
Stripe advanced error handling
Distinguishes error types and discusses uncertain outcomes. Useful for designing status investigation, bounded retries and accurate pending-versus-completed messages.
Making retries safe with idempotent APIs
Explains request identifiers and the design choices behind safe repeated operations. Apply the ideas with the guarantees of your actual provider.
OAuth 2.0 security best current practice
Technical guidance for securing delegated access. Useful when agents act through user-authorized integrations; identity and token checks still need action-level business rules.
Architecture articles and further reading
Read these for design trade-offs and provider-specific patterns, then test their relevance to your own system.
Building effective agents
Compares workflows and agents and explains common implementation patterns. Helpful when deciding whether a task needs autonomous action at all.
Advanced tool use on the Claude Developer Platform
Explores tool discovery and execution patterns. Use it to understand the capabilities you must constrain; more capable tool use is not the same as broader business authority.
Safety and security for AI agents
Discusses identity, authorization, guardrails, sandboxing and evaluation in ADK. Useful for mapping a written plan to several complementary implementation controls.
Turn reading into a reviewable plan
Choose one action and record its evidence, routine limits, approval route, prohibited behavior, failure handling and recovery owner. Ask your implementation team which system will enforce each boundary and how they will test it.
Build a permission plan or explore the Permission Plan Review service. Found a broken link or a useful original source? Send a resource suggestion.
