The Good
Pope Leo XIV used his first encyclical (an official pastoral letter written by the Pope) to share ideas on safeguarding humanity in the age of AI. The core position: technology is never neutral, and AI must be "disarmed" from the logic of military, economic, and cognitive competition. No algorithm, he wrote, can make war morally acceptable. In an unusual move, the Vatican invited Anthropic co-founder Chris Olah (a non-believer) to present alongside the Pope. The choice wasn't random: Anthropic is the company that refused to let the US military use its technology for autonomous weapons (covered in our 4 March drop). In the same week, Anthropic and the Gates Foundation committed $200 million to deploy AI across global health, education and economic mobility initiatives in underserved regions.
Comment: We've been saying AI should elevate humanity since day one. Turns out the Pope agrees.

The Bad
Five Eyes (Australia, the US, UK, Canada and New Zealand) cybersecurity agencies released their first joint guidance on agentic AI. It's 30 pages on what can go wrong when AI systems plan, decide and act without a human checking every step.
Comment: The guidance itself is good. What's bad is that it had to be written at all – and that it opens by saying agents are already operating in critical infrastructure and defence sectors. These are the agencies responsible for protecting national security – ASD, NSA, CISA – and they're essentially saying: this stuff is already out there, it's already being given too much access, and most organisations haven't caught up. We spent the intro to this drop explaining why agents are exciting – and we stand by that. But exciting and risky aren't mutually exclusive. The same capability that lets an agent draft your documents and manage your inbox is a problem if it's given too much access or no one's watching what it does. Use agents, just don't hand them the keys to everything on day one – especially if you don’t understand how they work. (Let us help you with that!)

& The Ugly
Last week, US federal prosecutors unsealed the first major charges under the Take It Down Act – a law making non-consensual AI-generated intimate imagery a criminal offence. Two men were charged with creating deepfake pornography of 140 named victims, with the content racking up nearly 3 million views. They face up to two years in prison. The law, signed in May 2025, now also requires platforms to remove non-consensual deepfake content within 48 hours of a valid takedown request. For context on the scale of the problem: earlier this year, a CCDH study found Elon Musk's Grok generated an estimated 3 million sexualised images in just 11 days – including 23,000 of children – before the feature was restricted. Three teenage girls have since filed a class action against xAI, with the first hearing set for 18 June. Malaysia and Indonesia blocked the platform outright.
Comment: For months we've been covering deepfake abuse in analogue and asking where the teeth are. Now there are arrests and platforms have a 48-hour clock. It's a solid start. But two years' prison for 140 victims and 3 million views doesn't exactly scream deterrence – and Australia still doesn't have equivalent federal legislation.

If you've ever asked a chatbot the same question twice and got two different answers, you've already found the limit of chatbots for business. They're great when the stakes are low and the question is one-off. But the moment you try to run something repeatable through one – a monthly competitor scan, an invoice check, a weekly CRM review – the cracks show. Different answer every time. No way to see what it actually looked at. No way to fix it when it gets something wrong.

There's a way out of this, and it's more accessible than most people think. Let’s call it a Mastermind and Minions architecture, and once you see how it works, you'll start spotting where it could solve problems in your own business.

Chatbots are undeniably fun for ad hoc tasks, but if you are trying to run daily, repeatable business processes, they are clunky and difficult to automate. If you are running a monthly competitor analysis or checking incoming invoices against contracts, you cannot rely on a magic eight-ball – you need a system you can run repeatedly, inspect and improve. One way to get there is a 'Mastermind and Minions' architecture.

The Vietnam Test

To see how this works, let's explore how we would build an agent that recommends a family holiday destination – like Vietnam – and explains the ‘why’ behind its choice. Instead of asking one massive AI model a broad question, we deploy 10 specialised ‘minions’, asking each a narrow question about Vietnam. One minion investigates if the weather is suitable during the upcoming school holidays, while another evaluates the country's safety for young children. 

Standard chatbots actually do this kind of sub-component breakdown internally, but our architecture forces that logic out into the open to create an audit trail. There are three key components: 

The Mastermind

The minions provide the narrow insights, but the mastermind acts as the executive decision maker. It reads what each minion wrote and synthesises the final recommendation for the user. The mastermind is an adjudicator. It weighs contradictions – for example, the ‘Budget Minion’ loves the cheap flights, but the ‘Weather Minion’ is flagging a monsoon. Because every minion produces a numerical rating, we can quickly identify exactly which minion swayed the mastermind's final decision. Using Claude models as an example, deploying 10 smaller Haiku models as minions costs a fraction of a cent for each query, saving the heavy lifting for the Opus mastermind model, which costs 2 cents. 

Tuning the machine

Because we can review the exact evidence bundle each minion read and the JSON rating it passed to the mastermind, we have an auditable and adjustable system. This allows us to rapidly check and improve the agent's logic. If the ‘Safety Minion’ incorrectly gives Vietnam a 1/10 risk rating due to motorcycle accidents, we don't have to scrap the whole system. We simply adjust the prompt to clarify that the travellers will only be using cars, or we change the research bundle the minion ingests. 

Unlike a standard chatbot that forgets your preferences the moment you close the browser window (assuming you’re not paying for a subscription and have configured you user settings), this architecture creates a compounding business asset. Once you fix the motorcycle rule, the agent never makes that mistake again. Your team's specific expertise becomes permanently baked into the code. This level of adjustment only makes sense if you are building a scalable, repeatable, commercial-grade agent. 

From holidays to the office 

Astoundingly, building this system in Claude Code with a few ‘paulminions’ is a weekend project, even for someone with no coding experience. 

Next week, you could arrive at the office to a custom dashboard delivering insights on CRM activity, tracking changes for a key client or recommending a new product formulation. Just as you tuned the ‘Holiday Minion’ to ignore motorcycles, you can seamlessly tune your ‘CRM Minion’ to ignore low-tier leads. 

When your team reads that CRM recommendation, they won’t just see a black-box directive. They will see clickable footnotes showing exactly which minion raised a red flag and the specific document that caused it. With this architecture, your AI's advice becomes both accurate and transparent, allowing you to examine exactly what each minion saw and decided. Ultimately, this level of forensic auditability is the standard of governance that executive and compliance teams need to green-light AI agents into a daily workflow.

So now what?

Even if you never build one of these yourself, knowing the architecture exists changes how you evaluate AI tools. When a vendor tells you their product “makes decisions” or an internal team pitches an AI workflow, the question is no longer, “Is it accurate?” It's: “Can you show me what each component saw and decided?” That's a governance posture you can carry into a meeting on Monday, regardless of whether you can code.

And if you do want to give it a go, Claude Code is the most accessible starting point – you describe what you want in plain English and it builds the scaffolding for you. Start with one minion doing one job. Get that working before you add a mastermind. Most people who give up on agents do so because they tried to build the whole thing on day one.

Paul Voutier is an international development specialist, startup mentor and robotics engineer with 15+ years at the intersection of technology, agriculture and emerging markets. He built Ambit Robotics, a Melbourne agtech company applying AI vision systems to tomato yield projection and harvest optimisation, and is a contributing author to ‘Intelligent Decarbonisation’, which examines how AI can help cut carbon emissions.

How good. You’ve made it this far. Thanks for joining us again and please feel free to hit us up with any questions or feedback. Otherwise, see you in the next drop.

Yours in humanity,