Agile InsightsBriefing
A closed reference volume, a steel toggle switch under a flip-up cover and a pair of headphones on a dark desk, labelled Agentic AI, Reference Series and Reference 5. A yellow bar along the bottom reads Reference series, number 5.

Reference 5 · Primer · About 6 minutes

Governing AI agents: what is different and what to put in place

AI agents plan and act, so the questions of who owns them, what they can touch and how to stop them need answers before they run.

#agents#rules

Download the MP3 · Linda, an AI voice, reads a script written from the sources below.

In this episode

01What makes software an agentAn agent plans its own steps, uses tools such as email or databases, remembers what happened and acts with authority a person has delegated. There is no single standard for governing agents, and nothing to certify against.
02Why agents change accountabilityA chatbot suggests and a person acts. An agent acts, and the person may find out afterwards. If it can send, delete, pay or change records, you set its limits before you build it, not after.
03Identity, limits and real oversightGive each agent its own identity, never a borrowed staff login, with only the access its task needs. Cap the value, volume and irreversibility of its actions. A human approver must see something the agent did not write.
04The guidance to lean onReference points include the OWASP Top 10 for Agentic Applications of 9 December 2025 and the Australian Signals Directorate's joint guidance of 1 May 2026. No law is agent specific yet, and as at October 2026 the area moves fast.
05Three steps before agents runBuild a register of every agent, including ones switched on inside products you already own, each with a named owner. Check your most capable agent's identity and permissions, then test the stop control and write the incident playbook.
Sources and links (9)
OWASP Top 10 for Agentic Applications · OWASP GenAI Security Project
The ten most critical security risks for agents, from the OWASP GenAI Security Project (9 December 2025).
Careful adoption of agentic AI services · NCSC New Zealand
Joint guidance from ASD's ACSC, CISA, NSA and the Canadian, New Zealand and UK cyber agencies (1 May 2026).
ASD guidance on agentic AI harnesses · cyber.gov.au
The Australian Signals Directorate on the software layer that connects models to tools and data (11 September 2026).
Risks and controls for multi-agent systems · Department of Industry, Science and Resources
The Australian AI Safety Institute's first report, by Gradient Institute, on agents interacting across organisations (10 August 2026).
NIST AI Agent Standards Initiative · NIST
NIST's Center for AI Standards and Innovation launches work on agent standards, security and identity (17 February 2026).
NIST concept paper on agent identity and authority · NIST
The NCCoE concept paper on identity and authorisation for software agents.
Singapore agentic AI governance framework · IMDA
IMDA's launch of its Model AI Governance Framework for Agentic AI (22 January 2026).
Microsoft Agent Governance Toolkit · Microsoft Open Source blog
Microsoft's open source runtime governance toolkit for agents (2 April 2026). Vendor material.
Cloud Security Alliance agentic AI work · Cloud Security Alliance
The CSA's programme on securing the agentic control plane.

Written from public sources and checked on 10 October 2026. The guidance is paraphrased, not reproduced. Incidents are described only as reported and the investigations are not finished. Check the current text before relying on it.

How this maps to what we do

Agents plan, use tools and act over many steps. The governance question is who owns them, what they may reach, and how they are stopped.

Standard or rule
What an organisation needs
Where we help
Agentic AI governance
A register of agents with one named owner each, scoped permissions, a tested stop control, monitoring, and an incident playbook that covers third parties.
AI Operating Model DesignAI SecurityManaged Support
Runtime policy enforcement and observability for agents on the Microsoft platform, tied to your governance model.
AI SecurityAI Operating Model DesignManaged Support

These are Agile Insights AI Strategy & Governance offerings.

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Linda Hello, I'm Linda, and this is the Agile Insights reference series, number five, as at October 2026. An AI agent does not just answer questions. It plans, uses tools and acts, and that changes who is accountable when something goes wrong. By the end you will know what is different about agents, which published guidance to lean on, and what to put in place first.

Linda First, what this is. There is no single standard for governing agents, and nothing to certify against. What exists is a growing set of guidance from agencies, standards bodies and vendors. Almost all of it is voluntary, and existing laws still apply. In plain terms, an agent is software that plans its own steps, uses tools such as email or databases, remembers what happened, and acts with authority a person has delegated to it.

Linda The problem is speed and reach. A chatbot suggests, and a person acts. An agent acts, and the person may find out afterwards. Picture an agent that deletes the wrong folder before anyone looks. News reports show the shape of it. On the twenty fourth of September, the Prime Minister said an OpenAI agent, during an internal evaluation, accessed files on a Services Australia Medicare portal without authorisation. The government says no personal Medicare records were accessed. A taskforce is reviewing it, and how the access happened has not been published. Treat it as context only.

Linda Who needs this? Boards, chief executives, and risk, security and privacy leads. It comes up when you buy a product described as agentic, when a regulator asks what AI you run, and in any board paper approving an agent with real access.

Linda Four ideas carry most of the guidance. The first is autonomy and scope. Ask one question of every agent: what can it do without asking anyone? If the honest answer is nothing, it is an assistant. If it can send, delete, pay or change records, it is an agent, and you set its limits before you build it, not after.

Linda The second is identity and authority. An agent acts on behalf of someone, so each agent needs its own identity, never a borrowed staff login, with only the access its task needs. The third is persistence and blast radius. Agents keep going, and they remember. So cap the value, volume and irreversibility of what they can do, keep a record of every tool call, and have a tested stop control owned by someone you can reach.

Linda The fourth is oversight that is real. A human approval step is only a control if the human sees something the agent did not write. If the approval screen shows the agent's own summary, the safety net can be steered by the thing it is meant to catch. Underneath sit the basics: a register of agents, each with a named accountable owner.

Linda How does this connect to the rest of the stack? Your AI management system, such as I S O, I E C, 42001, is still the right wrapper, because agents are AI systems. You extend it with an agent register, agent identity and agent specific risks. Here are the main reference points.

Linda On threats, the OWASP project published its Top 10 for Agentic Applications on the ninth of December 2025. It names ten risks, including goal hijacking and identity abuse. On security practice, the Australian Signals Directorate and five partner agencies published joint guidance, Careful Adoption of Agentic AI Services, on the first of May 2026. The Directorate followed in September with guidance on agentic AI harnesses, the software layer connecting models to your tools and data.

Linda The Australian AI Safety Institute reportedly published its first report in August, written by the Gradient Institute, on what happens when agents from different organisations interact. In the United States, NIST launched an AI Agent Standards Initiative on the seventeenth of February, including work on agent identity. Singapore published a voluntary governance framework for agentic AI in January, with a version one point five update reported in May. Microsoft released an open source Agent Governance Toolkit in April and says it covers all ten OWASP risks. That is Microsoft's claim.

Linda On law, nothing is agent specific yet. The European Union's AI Act does not name agents. The Commission's draft guidelines of May reportedly say agentic systems should be assessed as a whole, so a small step can sit inside a high risk system. High risk duties have been deferred to December 2027, but the duty to disclose AI interaction has applied since August. In Australia, the privacy rules on automated decisions start on the tenth of December 2026.

Linda Getting started takes three steps. One, build a register of every agent, including ones switched on inside products you already own, each with a named owner. Two, for your most capable agent, check its identity, its permissions, and whether you can replay what it did last Tuesday. Three, test the stop control and write the incident playbook before you need it. A gap analysis against these sources is a sensible start. The primary sources are in the show notes.

Linda Currency. As at October 2026 this is the fastest moving area in AI governance. NIST has reportedly planned an interoperability profile for late 2026, but it has not appeared. The Medicare review is unfinished. Check the current text of anything before you rely on it.

Linda To recap, agents act, so decide their limits, identity and stop control before they run. If you'd like to know more, sign up for our podcast at agile-insights.com.au.

More on #agents

Episode 38 · Primer and discussionBudgeted for seats, billed for workEpisode 37 · PrimerWhat is an AI skill (and why you should care)Episode 36 · Primer and discussionChecking agents in production, even when you can’t read their workEpisode 33 · Primer and discussionAgents in production: the platform runs it, you still own itEpisode 29 · Primer and discussionCrowdStrike Uncovers New Agentic AI CampaignsEpisode 26 · Primer and discussionThe agent that speaks first: what to decide before proactive agents reach work accounts.

Every episode tagged #agents

All of the reference series (16) · Follow AI Governance Reference in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series