Australia issues guidance for government use of agentic AI

Australia’s Digital Transformation Agency (DTA) has issued an agentic AI addendum to its AI Technical Standard, providing guidance for government agencies exploring, developing or deploying agentic AI systems. The document provides best-practice guidance for agencies exploring, developing, or using agentic AI and states that existing requirements in the AI technical standard remain applicable.

The addendum says agentic AI systems may autonomously plan tasks, coordinate work, and trigger actions in real-world contexts. The addendum notes that agentic AI could improve the responsiveness, efficiency and consistency of public services, particularly in high-volume administrative environments, while also introducing new risks related to oversight, control and system behaviour.

The guidance defines agentic AI as systems capable of perceiving and interpreting their environment, maintaining an internal state, reasoning about objectives and autonomously executing actions within defined permissions and constraints. Agencies are advised to implement human oversight, operational safeguards, continuous evaluation processes and mechanisms that allow systems to be rolled back when necessary.

The addendum sets out guidance across the AI lifecycle, including governance and safeguards, memory management, workflow design, secure data exchange, technology selection, evaluation, tool integration, monitoring, and decommissioning. It also calls for clear human accountability, human-in-the-loop or human-on-the-loop oversight, auditable decision records, and orchestration layers.

The guidance recommends ongoing monitoring of agent behaviour, tool usage, memory functions, operational costs, latency, authorisations and changes in the operating environment. The addendum also recommends centralised oversight mechanisms, referred to as ‘control towers’, and calls for the secure decommissioning of agentic AI resources, including agents, associated data, memory stores, tools and system logs.

Why does it matter?

Agentic AI represents a shift from AI systems that generate outputs in response to prompts to systems capable of planning, coordinating tasks and taking actions with limited human intervention. While these capabilities could improve efficiency and service delivery, they also create new governance, accountability and security challenges.

Australia’s guidance reflects growing international efforts to establish safeguards for increasingly autonomous AI systems. The emphasis on human oversight, auditability and lifecycle governance highlights concerns that public-sector AI deployments must remain transparent, controllable and accountable as the technology evolves.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!