Skip to content

1. Introduction

1.1 Purpose of the Standard

This standard establishes the principles and methodologies for developing and deploying agentic systems within the WOG context. Its primary purpose is to ensure a cohesive, responsible, and effective approach to building production ready agents, specifically by offering compliance recommendations upfront.

This standard expands on previous work, including the Agentic AI Primer and an exploration of existing Agent SDKs. It also integrates other resources, such as our Agentic Risk and Capability Framework and MCP Governance Framework1, to offer a comprehensive perspective.

1.2 Who this Standard is for

For readers looking to comply the development or deployment of agentic systems within GovTech, Chapter 2 provides a compliance decision flowchart. This is followed by the compliance table itself, and references to other applicable compliance frameworks.

For interested developers of agentic systems, Chapters 3-5 provide exposition on key principles, practices, and risks to guide design considerations to be put in place to ensure safety and accountability. While these are not compliance requirements to be followed per se, we hope these chapters serve as a useful guide for developers of agentic systems.

1.3 Scope of the Standard

1.3.1 Definition of AI Agents and Agentic AI Systems

For the purposes of this Standard, an AI agent (or agentic AI system) is an autonomous system built around an augmented large language model (LLM) that meets any of the following:

  • Operates through one or more orchestration loops that interpret the LLM's outputs
  • Able to utilise tools
  • Able to modify memory or the external environment

This Standard will be applicable as long as an AI Agent is part of such a workflow. Having a predefined or semi-predefined path helps with security and compliance considerations but does not preclude the system from being an agent.

A detailed explanation of this definition, including underlying concepts and examples, is provided in Appendix 7.1.  


  1. On Innersource, requires Techpass to access ↩