Standards Reference & Sources
Every pattern carries a standards mesh pointing to articles, principles, or controls it helps satisfy. Below is the consolidated reference set, the full bibliography, and a note on method.
This page lists the real-world rules, laws, and guidelines that AI agents need to follow, made by governments and safety groups around the world — like Singapore's IMDA, the European Union, and international safety standards. It also lists every book, article, and expert this catalogue's patterns are based on.
These rules matter because AI agents that act on their own can cause real problems if nobody is checking them, like making the wrong decision or creating a security risk. Every pattern in this catalogue points back to the specific rule or article it helps satisfy, so you can trace exactly why each pattern matters and where the idea came from.
Primary frameworks
Scope and relevance of each framework to agentic AI. Standards mesh mappings throughout this catalogue are indicative, not legal opinion.
| Framework | Scope & Relevance to Agentic AI |
|---|---|
| IMDA MGF (2026) | Singapore Model AI Governance Framework for Agentic AI. Four dimensions: (1) assess and bound risks upfront; (2) make humans meaningfully accountable; (3) implement technical controls and processes throughout the agent lifecycle; (4) enable end-user responsibility. World's first agentic-specific governance framework. |
| EU AI Act | Arts. 9 (risk management), 10 (data governance), 11 (technical documentation), 12 (record-keeping/logging), 13 (transparency), 14 (human oversight), 15 (accuracy/robustness/cybersecurity), 50 (transparency obligations), 72 (post-market monitoring). Many agentic use cases (credit scoring, employment screening) fall under "High-Risk". |
| NIST AI RMF 1.0 | Four core functions: Govern, Map, Measure, Manage. Supplemented by NIST AI 600-1 (Generative AI Profile) covering 12 risk categories. Outcome-oriented rather than prescriptive; pairs well with EU AI Act (obligations) and ISO 42001 (management system). |
| ISO/IEC 42001:2023 | AI Management System (AIMS) standard. Certifiable via accredited bodies per ISO/IEC 42006:2025. Paired with ISO/IEC 23894 (AI risk management). Covers data governance, transparency, performance evaluation, continual improvement. |
| MAS FEAT | Monetary Authority of Singapore: Fairness, Ethics, Accountability, Transparency principles. Applies to use of AI and data analytics in Singapore's financial sector. Complemented by MAS TRMG (Technology Risk Management Guidelines). |
| HKMA GL-1 | Hong Kong Monetary Authority: High-level Principles on Use of Generative AI by Authorized Institutions. Emphasis on governance, accountability, fair treatment of customers, data management, cybersecurity. |
| BNM RAI | Bank Negara Malaysia: Responsible AI principles. Covers soundness, accountability, fairness, data integrity, transparency for financial institutions. |
| OWASP LLM Top 10 | Top 10 threats for LLM applications: prompt injection, insecure output handling, training data poisoning, model denial of service, supply chain vulnerabilities, sensitive information disclosure, insecure plugin design, excessive agency, overreliance, model theft. OWASP AIVSS extends to agentic scoring. |
| Gartner TRiSM | AI Trust, Risk, and Security Management. Pillars: explainability/model monitoring, ModelOps, AI application security, privacy. Industry-standard framing for enterprise AI risk. |
| CSA SG Addendum | Cyber Security Agency of Singapore — Addendum on Securing Agentic AI (Oct 2025). Practical controls for agent system owners: identity, tool whitelisting, memory poisoning defence, sandboxing. |
| AAGMM | Agentic AI Governance Maturity Model (arxiv 2604.16338): 5 levels × 12 governance domains grounded in NIST AI RMF and ISO 42001. Includes sprawl taxonomy: functional duplication, shadow agents, orphaned agents, permission creep, unmonitored delegation chains. |
| GDPR · PDPA | Data protection. PDPA (Singapore) advisory guidelines on use of personal data in AI recommendation and decision systems (2024). GDPR Art. 22 (automated decision-making), data residency, purpose limitation. |
| BCBS 239 | Principles for effective risk data aggregation and risk reporting. Relevant to agents in risk and regulatory reporting flows — especially observability, lineage, adaptability, accuracy. |
Clause and domain breakdown
Where each framework's own internal structure — clauses, articles, governance domains — maps onto the catalogue.
Singapore IMDA Model Governance Framework for Agentic AI
World's first agentic-specific governance framework. Four dimensions: (1) Assess and bound risks upfront; (2) Meaningful human accountability; (3) Technical controls across lifecycle; (4) End-user responsibility.
- Risk Bounding at Design Time
- Agent Identity as First-Class Principal
- Use-Case Scoping with Prohibited Uses Register
- Meaningful Human Accountability
- Automation Bias Countermeasures
- Technical Controls Across Lifecycle
- Post-Deployment Monitoring
- Offline / Kill-Switch Mechanisms
- End-User Responsibility Patterns
EU AI Act — Article-Mapped Obligations
Regulation (EU) 2024/1689. High-Risk AI systems (many agentic BFSI uses) face structured obligations mapped to specific articles.
- Art. 9 — Risk Management System across lifecycle
- Art. 10 — Data Governance (training, validation, testing)
- Art. 11 — Technical Documentation (Annex IV)
- Art. 12 — Automatic Logging (min 6 months for high-risk)
- Art. 13 — Transparency to Users
- Art. 14 — Human Oversight (approval, stop, override)
- Art. 15 — Accuracy, Robustness, Cybersecurity
- Art. 50 — Transparency Obligations for GenAI
- Art. 72 — Post-Market Monitoring
- Art. 86 — Right to Explanation
NIST AI Risk Management Framework + AI 600-1 GenAI Profile
Four core functions — Govern, Map, Measure, Manage — supplemented by NIST AI 600-1 Generative AI Profile covering 12 risk categories.
- GOVERN 1.2, 2.1 — Policy mechanisms
- GOVERN 1.4 — Accountability structures
- GOVERN 1.6 — AI system inventory
- GOVERN 1.7 — Safe decommissioning
- GOVERN 3.2 — Oversight configurations
- MAP 1.1, 1.5 — Intended purpose and risk tolerances
- MEASURE 2.3, 2.5 — Testing and monitoring
- MEASURE 2.7 — Safety
- MEASURE 2.8, 2.9 — Logging, traceability, explainability
- MEASURE 2.13 — Human review
- MANAGE 1.2 — Resource allocation
- MANAGE 4.1, 4.2 — Third-party management
ISO/IEC 42001:2023 — AI Management System (AIMS)
Certifiable management-system standard via ISO/IEC 42006:2025 accreditation. Paired with ISO/IEC 23894 AI risk management.
- Clause 6.2 — AI objectives
- Clause 8 — Operational planning and control
- Clause 9 — Performance evaluation
- Clause 10 — Continual improvement
- Annex A — AIMS controls (including AI system inventory A.6)
OWASP Top 10 for LLM Applications + AIVSS
Top threat categories for LLM and agentic applications. AIVSS extends to agent-specific scoring.
- LLM01 — Prompt Injection (direct and indirect)
- LLM02 — Insecure Output Handling
- LLM03 — Training / Memory Poisoning
- LLM04 — Model Denial of Service
- LLM05 — Supply Chain Vulnerabilities
- LLM06 — Sensitive Information Disclosure
- LLM07 — Insecure Plugin / Tool Design
- LLM08 — Excessive Agency
- LLM09 — Overreliance
- LLM10 — Model Theft
Cyber Security Agency of Singapore — Addendum on Securing Agentic AI
Practical security controls specific to agent system owners.
- Agent Identity controls
- Agent Sandboxing
- Tool Allowlisting
- Indirect Injection Defence (agent-specific)
- Memory Poisoning Controls
- Least-Privilege Tool Credentials
Monetary Authority of Singapore — FEAT Principles
Fairness, Ethics, Accountability, Transparency — guidance for use of AI and data analytics in Singapore's financial sector. Complemented by MAS TRMG.
- Fairness — fair treatment of customer segments
- Ethics — responsible use of AIDA
- Accountability — internal governance and review
- Transparency — explainable outcomes to customers
- MAS TRMG — Technology Risk Management Guidelines
Hong Kong Monetary Authority — High-Level Principles on GenAI
Principles for authorised institutions using Generative AI.
- Governance and accountability
- Fair treatment of customers
- Data management
- Cybersecurity and model risk
Bank Negara Malaysia — Responsible AI Principles
Responsible AI principles for Malaysian financial institutions.
- Soundness
- Accountability
- Fairness
- Data integrity
- Transparency
Agentic AI Governance Maturity Model
arXiv 2604.16338 — 5 maturity levels × 12 governance domains. Defines agent sprawl taxonomy.
- GD-01 — Lifecycle & Identity
- GD-02 — Identity Management
- GD-03 — Permissions & Least Privilege
- GD-04 — Policy-as-Code
- GD-05 — Delegation
- GD-06 — Authority
- GD-07 — Sprawl Scanning (duplicates · shadow · orphaned · creep · unmonitored delegation)
- GD-08 — Audit
- GD-09 — Monitoring
- GD-10 — Decommissioning
- GD-11 — Incident Response
- GD-12 — Governance Domains Coverage
Gartner AI Trust, Risk, and Security Management
Industry-standard framing for enterprise AI risk across four pillars.
- Explainability / Model Monitoring
- ModelOps
- AI Application Security
- Privacy
Data Protection Regulations
EU GDPR and Singapore PDPA including 2024 advisory guidelines on personal data in AI systems.
- GDPR Art. 5 — Principles (minimisation, storage limitation)
- GDPR Art. 17 — Right to erasure
- GDPR Art. 22 — Automated decision-making
- GDPR Art. 32 — Security of processing
- PDPA Singapore — Protection obligation
- PDPA Singapore — Advisory Guidelines on AI (2024)
BCBS 239 — Risk Data Aggregation and Reporting
Basel Committee principles especially relevant to agents operating in BFSI risk and reporting flows.
- P2 — Data architecture and IT infrastructure
- P3 — Accuracy and integrity
- P4 — Completeness
- P5 — Timeliness
- P6 — Adaptability
- P7 — Accuracy of reporting
- P11 — Distribution
Every pattern traced
Industry, academic, and vendor sources (S1–S10) plus regulatory and standards references (R1–R12) synthesised throughout this catalogue.
Industry & Academic Sources
- S1Anthropic — *Building Effective Agents* (Schluntz & Ingham, Dec 2024) and *Building Effective AI Agents: Architecture Patterns and Implementation Frameworks* (2025).
- S2Antonio Gulli (Google, Office of CTO) — *Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems* (Springer, Dec 2025, 424 pp, 21 patterns).
- S3Dao et al. — *Agentic Design Patterns: A System-Theoretic Framework* (arXiv 2601.19752, Jan 2026).
- S4*Agentic AI: Architectures, Taxonomies, and Evaluation of LLM Agents* (arXiv 2601.12560).
- S5AWS Prescriptive Guidance — *Foundations of Agentic AI on AWS*.
- S6Google Cloud Architecture Center.
- S7Microsoft Agent Factory.
- S8Salesforce Enterprise Agentic Architecture.
- S9LangChain / LangGraph.
- S10Hugging Face Agents Course.
Regulatory & Standards References
- R1IMDA Model Governance Framework (MGF) for Agentic AI.
- R2CSA SG Addendum — Cyber Security Agency of Singapore, Addendum on Securing Agentic AI.
- R3EU AI Act.
- R4NIST AI RMF (AI Risk Management Framework).
- R5ISO/IEC 42001 / 23894.
- R6MAS FEAT — Monetary Authority of Singapore Fairness, Ethics, Accountability, Transparency principles.
- R7HKMA GL-1 — Hong Kong Monetary Authority High-Level Principles on Generative AI.
- R8BNM RAI — Bank Negara Malaysia Responsible AI principles.
- R9OWASP LLM Top 10.
- R10Gartner TRiSM — AI Trust, Risk, and Security Management.
- R11AAGMM — Agentic AI Governance Maturity Model.
- R12GDPR / PDPA — EU General Data Protection Regulation and Singapore Personal Data Protection Act.
Why a pattern catalogue
Software architecture matured once its patterns were codified. Data architecture is following the same arc. Agentic AI is at the earliest phase — patterns are described variously across blog posts, vendor docs, one book, and a handful of arXiv pre-prints.
This catalogue treats each design decision as a reusable primitive — a named solution to a recurring problem in a specific context — rather than as a framework feature. Reflection is a pattern; LangGraph is an implementation.
Every pattern has three depths: a sketch (card), a reference (modal body), and a mapping (standards grid). Start shallow. Deepen only when implementing.
Honest Caveats
The agentic AI field is three years old. Unlike BCBS 239 (2013) or DAMA-DMBOK, there is no ISO-grade canonical pattern list yet. This catalogue synthesises the authoritative sources named above; it is not a standard.
Gulli's 21-pattern corpus is the closest to a standard — but it is one author's synthesis, not a committee standard. Anthropic's six workflow patterns are the most widely replicated.
The IMDA Model Governance Framework for Agentic AI (22 Jan 2026) is the world's first agentic-specific governance framework. Singapore is ahead of the regulatory curve.
Standards mesh mappings are indicative. They do not constitute legal advice or regulatory endorsement.
Every pattern in this catalogue is traceable to at least one public reference. Where the field lacks a widely recognised name, the more general name is preferred. No proprietary methodology is reproduced.