Trustworthy AI agents: New technical guidance

Published as an addendum to the AI technical standard, the new artificial intelligence (AI) addition sets clear expectations for how agencies manage and operate agentic AI systems. Agentic systems operate autonomously to achieve specific goals with minimal human supervision.

It builds on the existing AI technical standard and complements the Policy for the responsible use of AI in government. 

Together, they form a broader suite of guidance that supports the safe, transparent and accountable use of AI across government to deliver better services for all people and businesses. 

Supporting the next generation of AI in government

Agentic AI systems can autonomously plan tasks, coordinate actions, and interact with tools or services to achieve defined outcomes within set guardrails that are overseen by humans.

These systems represent a shift from traditional AI tools that primarily analyse data or generate content, to systems that can autonomously plan, reason and act on complex tasks with minimal human involvement.

For government, this can support improved service delivery, faster responses to complex operational needs and reduced administrative burden, particularly in high-volume, rules-based workflows.

“As these systems move from experimentation to operational use, their increased autonomy introduces new considerations for agencies across government,” explains Lucy Poole, Deputy CEO for Strategy, Planning and Performance at the DTA.

“This underscores the need for clear, practical and consistent guidance tailored to the government context, helping agencies adopt these technologies safely, responsibly and with confidence.”

Strengthening safeguards and accountability

The addition sets out a best practice approach to agentic AI, recognising that different use cases require different levels of control and oversight. 

It provides guidance across the full lifecycle of agentic AI systems, including:

• governance, controls and system design
• data and memory management
• evaluation against defined outcomes
• human oversight and monitoring
• mechanisms to intervene or safely shut down systems

Human oversight maintains accountability for outcomes and helps ensure systems remain safe and aligned with their intended purposes, even where they operate with greater autonomy.

“Agentic AI solutions have the potential to establish automated workflows that transform the way governments staff deliver services, and – by extension – how those services deliver value for the people and businesses that need them,” emphasises Ms Poole. 

“From pulling all the necessary data required for comprehensive compliance reviews, to collating all the necessary information for staff to efficiently approve service applications, we look to balance the solutions capacity to reduce friction with services to near zero while maintaining appropriate safeguards and protecting both users and agencies.”

“This guidance helps government realise the benefits of more autonomous systems while maintaining clear accountability, oversight and appropriate human control.”

Maintaining human oversight

A core feature of the addition is its focus on human oversight, safety and accountability as these AI systems take on more complex tasks autonomously. This ensures systems operate safely within defined parameters while remaining under the continuous supervision of authorised human personnel.

This enables authorised personnel to:

• monitor system behaviour in real time
• intervene or adjust actions where needed
• pause or stop system activity if risks emerge

Next steps

As agentic AI capabilities continue to evolve, this addition provides a clear and consistent foundation for agencies to adopt more autonomous systems safely across government. It unlocks new capabilities for public servants while keeping the interests of Australians at the centre of decision-making.

It strengthens how agencies use agentic AI with appropriate safeguards in place, ensuring systems operate with robust oversight and accountability and in line with community expectations.

The new agentic AI addendum to the AI technical standard is available on digital.gov.au.