The lead: Australia now has a plain-language guide to AI agents
On 26 October the National AI Centre published Understanding and using agentic AI, a free guide for small and medium-sized enterprises that are considering, buying, implementing or already using AI agents. It is written for business owners, managers and staff without specialist expertise, and it builds on the Centre's existing Guidance for AI Adoption. The Centre says it was developed after consultation with 175 representatives from industry, academia, community and government, and that it draws on research from the Australian AI Safety Institute.
The guide defines agentic AI as AI that can plan, choose steps, use tools and take multiple actions towards a goal, with some ability to act on its own. Its simplest test is worth remembering: ask who takes the next step. If the tool returns information or suggestions and a person decides what happens next, it is usually not agentic. If it can act towards a goal by using tools and making decisions, it probably is. The guide also points out that agentic features can arrive inside software a business already uses, without being labelled as an agent, such as a booking system that schedules appointments and sends reminders on its own.
The practical core is four checks to complete before an agent is switched on:
- Access: give the agent only the information, systems and tools it needs for the task.
- People: decide when a person must review, approve or stop an action, particularly where customers, staff, finances or important processes are affected.
- Visibility: make sure you can see what information the agent used, what it did and what it changed.
- Stop: know how to pause it, remove its access, halt its actions and recover from mistakes before it goes live.
Two passages stand out for anyone in a governance or compliance role. First, the guide says human oversight is not always enough, because an agent can work at a speed or scale that makes reviewing every action impractical, and review only works when the reviewer has the time, information, expertise and authority to intervene. Second, it lists situations where a business should pause and get cyber security, privacy, legal or technical advice first, including agents that can reach bank accounts and payment systems, passwords, staff or customer records, health or legal information, emails, orders and refunds, or anything that publishes to a website or social media account.
It also restates a point that is easy to lose in vendor material: existing laws apply when AI is used, and organisations using AI remain responsible for identifying and responding to harm. The guide recommends starting with one small, clearly defined task with restricted access, and it says plainly that starting small limits the consequences of a problem without removing the risk.
GitHub makes model routing the default choice it can be
GitHub announced on 28 October that Copilot Fusion is generally available. Instead of a developer choosing a model for each task, Fusion decides which models to use for each prompt and how they work together, at one of three effort levels: Standard, High or Max. It can answer with a single model, let a cheaper model draft and escalate to a stronger one if a quality check fails, or have one model draft while another reviews.
The governance detail matters more than the routing. Fusion only uses models enabled in the enterprise and organisation model policies, and administrators can remove a model from Fusion without removing it for everyone. Billing is per model used, at each model's standard rate, with a limited-time 30% discount on paid plans. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. That is a vendor claim. Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. GitHub states no regional limit; Australian availability is unconfirmed beyond that.
For teams with a model policy, the useful question is now which models the router is allowed to reach, not which model each developer picked.
OpenAI attaches API credits to some ChatGPT plans
OpenAI's ChatGPT release notes for 28 October announced promotional API credits for eligible Pro and Business subscribers. The help article sets the amounts in US dollars: US$100, US$200 or US$500 a month on the three Pro tiers, and for Business, US$20 per paid Standard seat and US$100 per paid Premium seat, pooled and capped at US$500 a month per workspace. Plus plans and free trials do not qualify, Enterprise and Edu are not listed as eligible, and the paid subscription must have been active for at least seven days before claiming.
The controls deserve attention before anyone claims. One subscription links to one API organisation, the person activating must own both the ChatGPT account or workspace and the API organisation, monthly credits expire at the end of the billing cycle without rolling over, and the credits do not cover Codex usage through API keys. ChatGPT and API billing stay separate. OpenAI says availability also depends on the enabled cohort and country, so Australian access is unconfirmed until it appears in an account.
A benchmark rebuilds itself around private tests
Artificial Analysis, a benchmarking firm, released Intelligence Index v5 on 27 October. The firm says 60% of the index now rests on evaluations with private test sets or held-out answers, which reduces the chance that a model has seen the questions in training. The index weights knowledge work, coding, science and general knowledge at 25% each. On its first results, Claude Opus 5.5 led at 58, ahead of GPT-6 Astra at 57, and no single model led all nine evaluations. Its cost finding is the more useful one for buyers: among the ten highest-scoring models shown, cost per index task ranged from US$0.96 to US$13.04, more than 13 times.
A composite score is one firm's weighting of one set of tasks. It is a reason to shortlist and test, not a result to plan around.
Also in the window
ChatGPT's release notes for 27 October added iCloud Mail and Calendar plugins, so iCloud users can connect mail and calendar to search email, check schedules and update events, on all ChatGPT plans on the web, desktop app and iOS. The notes give no regional statement. Connecting a personal mailbox to any assistant is a decision to make deliberately, not by default, particularly where work email is forwarded to a personal account.
For how these items fit with the rest of the week, see the weekly review for 26 October to 1 November, and for every edition, AI Updates.
One thing to test this week
Take one tool in your workplace that already acts on its own, such as a booking system, an inbox rule or an AI feature that sends or updates things, and run the National AI Centre's four checks on it in under 30 minutes. Write one line for each. What can it access, and does it need all of it? Who approves or stops its actions? Where is the record of what it did? How would you switch it off today? Any line you cannot fill in is your first piece of work.
General information only. Not legal, compliance, financial, or professional advice.
TheAICommand. Intelligence, At Your Command.
