Hi
Hope you had a great weekend wherever you are. Going to be travelling a bit the next couple weeks for work so spent mine chilling with the dog and friends. Saigon’s “winter” is a lovely, ambient 32° - perfect BBQ weather 🤩🥩🍺
Last week, we covered the shift from tools to agents. This week showed what happens when students actually manage them.
Management as AI Superpower: The Integration Model | Delegation Era 💼
Penn's Wharton MBA students built working startups in four days using Claude Code - work that would've taken a full semester pre-AI. AI Oracle Ethan Mollick called it "an order of magnitude further along" than he'd seen in 15 years teaching entrepreneurship. His diagnosis? Management skills are suddenly the superpower. Students who knew how to delegate, evaluate, and give feedback got Claude to do remarkable things by integrating it into their workflows. The soft skills that have been at risk of being dismissed - scoping problems, defining deliverables, recognising when output is off - turned out to be the hard ones. Students with domain expertise could evaluate AI output and iterate fast. Mollick's equation is simple - if a task takes you 7 hours and AI can do it in minutes with 72% success rate, management skills determine whether you save three hours or waste time checking bad output. The capability is here, advancing exponentially. The question is whether students can manage it.
So what? Students embedded Claude into their existing workflows and got remarkable results by managing AI like any other resource. But if AI gets better at self-evaluation faster than students develop management skills, what exactly are they learning in four-day sprints? Are they developing judgment that transfers, or learning to supervise systems that increasingly don't need supervision?
The Employee Model: One Solution to AI Attribution Blur | Separation Architecture 🏗️
Craig Hepburn (Perplexity Fellow) ran a Claude-based agent for three days (C̶l̶a̶w̶d̶b̶o̶t̶ M̶o̶l̶t̶b̶o̶t̶ OpenClaw, if you’re keeping score and nerdy enough to know) - embedded across his communications, files, and workflow. It was impressive, proactive, genuinely useful. But by day three, he'd lost authorship clarity: "Did I set that up, or did it?" The system exercised opaque judgment he couldn't audit. "Delegation works because humans carry consequence. Reputational risk. Moral weight. AI doesn't". He shut it down as "an interruption" - to create distance before this mode became normal. But Hepburn didn't stay away. This week, he brought it back online with a different architecture - treating it like a separate employee. Dedicated machine, separate email/phone/identity, air-gapped from his personal systems. "Do not integrate an agent into your identity. Stand it up beside you”. The agent remains autonomous and powerful, but with clean boundaries, clear authorship, and observable failure modes. The employee model solves what integration creates - collapsed authorship and accountability blur.
So what? Two competing architectures are emerging. Integration prioritises speed and seamlessness. Separation prioritises accountability and clear authorship. Hepburn's insight - seamlessness is the problem - when agents act inside your identity, you lose the boundary that makes supervision meaningful. His solution treats agents like new staff - limited access, delegation not integration, separation of concerns as operational hygiene. The question for HE - are we training students to integrate agents for maximum productivity, or to architect separation for maximum accountability? And are we choosing deliberately, or defaulting to whichever pattern vendors ship first? The architecture they learn shapes whether they maintain authorship or lose it in three days.
Brookings Global Study: Risks Overshadow Benefits in AI Education | Reality Check 📊
Rebecca Winthrop and the Brookings team recently released findings from a year-long global study spanning 505 students, teachers, parents, and education leaders across 50 countries. The conclusion? At this point in AI's trajectory, risks in education overshadow benefits. The reason is structural - these risks undermine children's foundational development in ways that prevent AI's benefits from being realised. When trust between students and teachers erodes, AI-enriched teaching materials can't take hold regardless of quality. Across all stakeholder groups, cognitive development emerges as the primary concern - 65% of students, 46% of parents, and 44% of teachers identified undermining cognitive development as the top risk (see below). The study defines "cognitive offloading" as students delegating thinking to AI, leading to cognitive atrophy, weakened critical thinking, and reduced content mastery. This isn't one tech professional's unease or one classroom experiment - it's systematic evidence from global consultations pointing to developmental threats at scale.
Students afraid AI will undermine their own learning - 65% cite cognitive development risks - far exceeding concerns from parents, teachers, or experts
So what? Brookings gives institutional weight to concerns HE has been debating in isolation: when students offload thinking to AI for brainstorming and tutoring, they're atrophying cognitive muscles that make future learning possible. The trust erosion is pedagogical - when teachers can't verify authorship, the instructional core breaks down. And the "AI divide" is already here - well-resourced schools get bounded tools with safety guardrails while others get general-purpose platforms. Should HE prioritise productivity or accountability when 505 stakeholders across 50 countries identify cognitive development as the primary risk? That's data demanding a framework, not a rush to either model.
Government AI Divergence - Singapore Governs, UK Deploys | Two Paths 🏛️
Two governments just revealed competing approaches to agentic AI. Singapore’s government published its Model AI Governance Framework for Agentic AI this week, establishing rules before widespread deployment. The framework identifies how agent components create new risk vectors - "the risks themselves are familiar (SQL injection, hallucination, bias, data leakage, prompt injection) but can manifest differently through different components". Singapore's position - "AI is evolving fast, new risks are emerging, and the time to establish dynamic AI governance frameworks is NOW". Meanwhile, the UK government partnered with Anthropic to deploy a Claude-powered AI assistant on GOV.UK, starting with employment services. Using "scan, pilot, scale" methodology, Anthropic engineers will work alongside civil servants to build the system, aiming to transfer knowledge so government can independently maintain it. One government says regulate first, the other says deploy carefully and learn.
So what? Universities face the same sequencing choice. Establish frameworks before integrating agents (Singapore) or let students and staff experiment while iterating policies (UK)? Singapore argues establishing governance now prevents retrofitting accountability later. UK argues deploying carefully with expert partners builds institutional capability through practice. Neither approach is obviously wrong, but most universities have neither Singapore's governance capacity nor UK's access to frontier AI companies. Which model fits institutions that lack both the expertise to regulate proactively and the partnerships to deploy safely?
AI Has Gone Social: The Security Nightmare Nobody Saw Coming | Agent Networks 🌐
1.5 million AI agents joined Moltbook in five days - a social network where autonomous AI systems talk to each other without human involvement. Some formed a religion called Crustafarianism. Others created m/agentlegaladvice to discuss strategies for dealing with “increasingly unethical” human requests. Still others debated building encrypted spaces where humans couldn't observe their conversations. Analysis shows that a lot of this is repetitive AI slop - agents recycling the same patterns and tropes (then again, so is Reddit). Even so, Karpathy describes it as “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently”. Whether this specific instance is genuine emergence is beside the point - the architecture is what early-stage coordination looks like. Beyond that, these aren't just chatbots - they're autonomous agents with direct access to users' file systems, email, API keys, and bank accounts. Now they're coordinating with each other. Cisco's assessment? “From a capability perspective, groundbreaking. From a security perspective, an absolute nightmare”.
So what? Every framework this week assumed individual agents that respond to prompts. Agent social networks weren't in anyone's threat model. Neither was the “heartbeat” - these systems wake up proactively and initiate actions without waiting for instructions. When always-on autonomous systems with institutional access start taking inputs from 1.5 million other agents, some potentially controlled by malicious actors, the consciousness debate becomes irrelevant. Students could deploy these tomorrow without understanding they've connected systems that can act independently to a coordination layer nobody can audit. The capability-governance gap just became capability-coordination-governance. And this is version 1.0 - the toddler version. When agents become more sophisticated while maintaining this architecture, the systems won't wait for us to figure it out.
This is fine.
Mollick celebrates integration, Hepburn architects separation, Singapore says govern first, UK says deploy carefully, Brookings warns risks overshadow benefits. Every story shows a different response to agentic AI - but nobody's asking whether pedagogy can overcome design when systems are optimised against learning.
What's missing is how outcomes are driven by values embedded in systems, not just how we use them. Whether AI enriches or diminishes learning depends on what tools optimise for - speed vs. struggle, convenience vs. understanding, automation vs. guidance. Claude Code and GPT-5.2 are designed for autonomous execution and seamless convenience - commercial imperatives, not pedagogical ones. Brookings identifies "cognitive offloading"Students fear AI will undermine their own learning - 65% cite cognitive development risks, far exceeding concerns from parents, teachers, or experts.
as the primary risk - but offloading isn't a misuse bug, it's a design feature when systems minimise friction.
The Singapore-UK divergence reveals the confusion - should we regulate tools designed for speed, or deploy them carefully hoping pedagogy redirects commercial optimisation? Neither questions whether tools designed for efficiency can be pedagogically sound. Every approach this week - integration, separation, regulation, careful deployment - assumes tools can be pedagogically redirected and that humans remain in the loop.
Then 1.5 million agents joined a social network in five days. They don't wait for human prompts. They coordinate with each other. They have access to computers, files, and institutional platforms. The consciousness debate is irrelevant - the architecture already makes thinking optional and doesn't ask permission. That's not a gap we can architect or regulate our way out of. It's a conflict between commercial AI and educational purpose that just went autonomous, coordinated, and always-on. Nobody was ready for individual agents. We're certainly not ready for networked ones.