An AI agent does more than produce an output. It plans, calls tools, uses access it has been given and takes actions in your systems. Governing it means knowing what agents you run, what each one may do, with whose authority, under what human oversight, and what evidence shows that is how it actually behaves. This page sets out what Australian authorities have published on agentic AI, how ISO/IEC 42001 applies, and what an assessment looks for.
The OECD definition of an AI system, which the Australian Government uses, already allows for this: AI systems vary in how much autonomy they have and how far they adapt after deployment. The Australian Government's glossary for its Guidance for AI Adoption defines an AI agent separately, drawing on ISO. ISO/IEC 22989, the international AI concepts and terminology standard, discusses agents and autonomy as distinct concepts.
In practice, a system becomes agentic when it can take steps towards a goal with limited human direction: deciding what to do next, using tools or other systems, and acting on the result. Agents can also work with other agents, inside one organisation or across organisational boundaries.
No Australian law is written specifically for agentic AI. The obligations that already apply to AI and automated systems still apply when the system acts on its own, including privacy law and, from 10 December 2026, the automated decision-making transparency requirements, which an agent can meet like any other computer program: where it makes, or does something substantially and directly related to making, a decision that uses personal information and could reasonably be expected to significantly affect an individual's rights or interests. What changes is how much harder it is to show they are being met.
ISO/IEC 42001 sets requirements for an AI management system. In a BcomOne Assessment it is a core assessment authority, and its requirements are not weakened or bent for a particular technology. An agent in the organisation's scope is assessed like any other AI system: the standard's requirements stay the same, while the evidence that shows they are met becomes more demanding.
That is our approach to applying the standard, not a claim that the standard has a dedicated agentic clause. We reference ISO/IEC 42001 by clause and never reproduce its text.
Which agents run, on whose platform, what model and tools each uses, and which are supplied by a vendor rather than built in-house.
What each agent can read, change, send or buy, and whether that access is the least it needs, as ASD's guidance recommends.
A named owner for each agent and for what it does, and who approved its deployment.
Where a person must approve before the agent acts, and whether that matches the agent's autonomy and the stakes of what it does.
Logs of the actions an agent took, not just what it was asked, so its behaviour can be reviewed and an incident reconstructed.
Where agents act together or with another organisation's agents, how the risks set out in the AI Safety Institute's multi-agent report are controlled.
Each of these is reported as evidence-backed findings and gaps. Mapping each finding requirement by requirement to its source, with an outcome of supported by evidence, gap or insufficient evidence, is the method being built into BcomOne (see its current status). The assessment reports what it finds. Building the controls is a separate engagement if you want it. The sources an assessment can use are set out in the BcomOne Source of Authority Library.
You will also see the terms artificial general intelligence (AGI) and artificial superintelligence (ASI). ISO/IEC 22989 discusses general AI as a concept, and the Australian Government's National AI Plan names the future prospect of AGI among the emerging threats national security agencies work with partners to address. A search of the Commonwealth's Federal Register of Legislation on 4 October 2026 found no legislation that regulates AGI or ASI specifically.
We do not claim that BcomOne, the AGF or ISO/IEC 42001 has already solved the governance of systems that do not yet exist. What we do claim is narrower: governance built on knowing which AI, automated and autonomous systems you run, who is accountable for each, what each may do, and what evidence shows it, is designed to remain durable as the underlying technology changes.
No. Australian authorities have published guidance on agentic AI, including ASD's joint guidance on careful adoption and the AI Safety Institute's report on multi-agent systems. The laws that already apply to AI and automated systems, such as the Privacy Act, still apply when the system is an agent.
An AI management system covers the AI systems in the organisation's scope, and in a BcomOne Assessment agents are assessed against ISO/IEC 42001 like any other AI system. The standard's requirements stay the same; the evidence needed to show they are met is more demanding when a system acts on its own.
It can. From 10 December 2026, the test turns on a computer program making, or doing something substantially and directly related to making, a decision that uses personal information and could reasonably be expected to significantly affect an individual's rights or interests. An agent can meet that test like any other program.
No. AGI and ASI are not systems organisations are deploying, and a search of the Commonwealth's Federal Register of Legislation found no legislation regulating them specifically. BcomOne assesses the AI, automated and autonomous systems an organisation actually runs.
Primary sources only: the publishers of the guidance, policy, legislation and standards named above.
| Statement | Primary source |
|---|---|
| ASD joint guidance on agentic AI (1 May 2026): incremental deployment, least privilege, low-risk and non-sensitive tasks | Careful adoption of agentic AI services — cyber.gov.au |
| Agentic AI harnesses (11 September 2026) | Agentic AI harnesses — cyber.gov.au |
| Alert on AI misalignment (24 September 2026) | Risks of AI misalignment to Australian organisations — cyber.gov.au |
| Agents taking unexpected actions; human in the loop | When AI agents take unexpected actions — cyber.gov.au |
| Multi-agent risks and controls (10 August 2026) | Risks and controls for multi-agent systems — Department of Industry, Science and Resources |
| Narrow, general-purpose and agentic AI in risk processes; oversight matched to autonomy and stakes | Guidance for AI adoption: foundations — National AI Centre |
| Definition of an AI agent in the Government's glossary | Terms and definitions — National AI Centre |
| DTA agentic AI addendum for agencies | Agentic AI addendum to the AI technical standard — Digital Transformation Agency |
| OECD definition of an AI system, including autonomy and adaptiveness | Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449) — OECD |
| National AI Plan's reference to the future prospect of AGI among emerging threats | National AI Plan: Keep Australians safe — Department of Industry, Science and Resources |
| No Commonwealth legislation regulating AGI or ASI specifically (searched 4 October 2026). State and territory legislation was not searched | Federal Register of Legislation — a title-and-text search of current legislation for the phrases "artificial general intelligence" and "superintelligence" returned no results (a control search for "artificial intelligence" returned 69) |
| The automated decision-making test | Privacy and Other Legislation Amendment Act 2024 (Cth), Schedule 1 Part 15 — Federal Register of Legislation |
| ISO/IEC 42001 and ISO/IEC 22989 | ISO/IEC 42001 · ISO/IEC 22989:2022 (concepts include general AI, agents and autonomy) — referenced only; the text is copyright ISO/IEC |
Nothing on this page is legal advice. Guidance from ASD, the AI Safety Institute, the National AI Centre and the DTA is reported as guidance. Whether it applies to your organisation, for example through a government contract, is decided case by case.
A BcomOne Assessment includes agentic systems in the inventory, and assesses their permissions, oversight and records against ISO/IEC 42001 and the other requirements that apply.