From Army Barracks to Berlin Classrooms: What AI Agents Get Wrong (and Right) About Real Learning

Sunday Reflection — 2026-06-18

Last week I watched an AI-powered “study buddy” stall out at Masterschool Berlin. Twenty-four adults, mostly mid-career, trying to shift into tech. The agent had all the right features: quizzes, code reviews, instant feedback, polite nudges if you missed a deadline. Two weeks in, the group chat had filled up with yawns and small jokes about “Clippy, but with a PhD.” Engagement flatlined. One guy muted the assistant and never turned it back on.

This wasn’t for lack of technical spark. Most of these people had taught themselves a career’s worth already, in German bureaucracy or medicine or finance. But the agent expected them all to learn the same way—with a step-by-step, boxed-up approach to Python that reminded me more of a compliance manual than anything human. For people juggling jobs and families, it felt insulting to be nagged about progress milestones with no sense of real context. What the designers forgot: grown adults don’t just want tips or reminders. They want to argue, to adapt, to poke holes in the material and force it to prove itself. The AI couldn’t answer a “Why does it matter?” question without retreating to a Wikipedia-tier explanation. It missed the mess—the negotiation, the peer debate, the ritual griping about group projects that’s half the point of learning together.

Thirty years ago in the British Army, the gap between what the manual said and what the job demanded was even wider. Anyone can recite Section 5 on bridge building. What counted was how you handled the plan falling apart on the riverbank, forty minutes behind schedule, with your best driver sick and a thunderstorm inbound. Every sergeant I knew could quote procedures, but excellence meant reading people, spotting exhaustion, and adjusting the plan on the fly. Any “instructional tool”—AI, army or otherwise—that doesn’t allow for that level of improvisation won’t earn respect. You can’t automate trust.

Compare this to the push for Agile in software. Proper Agile isn’t about a script, it’s about adaptation in the presence of mess. You do short loops. You reflect on what broke. You rewrite your process every sprint. Most AI learning agents still treat knowledge transfer as a conveyor belt. The tools hand out self-quiz prompts and flag you if you stray from the curriculum, as if becoming competent were the same as clearing a checklist. Real project work—that improvisational scramble to deliver under half the information you need—barely gets a look in.

Much of this tracks with the national character. German teams I’ve managed treat process documentation as sacred, sometimes to the point where the document eclipses the thing itself. An AI assistant that tracks every lesson, links to every regulation and best practice, feels reassuring here. But even Germans will ignore the agent if it starts tripping over specifics or if, heaven forbid, it wades into company culture and tries to prescribe behaviour. In the UK, the suspicion kicks in faster. Brits will smile and thank the robot, then ask each other in the hallway if anyone actually checked the checklist. More than once in London I’ve seen a “recommended learning path” get the polite nod and immediate burial.

The lesson: most AI agents are great at delivering a plan, lousy at what actually happens next. They remind me of the engineer’s joke: in theory, everyone should salute the plan; in practice, you salute the ones who rescue it when the wheels come off. No algorithm knows when the apprentice in the back is quietly three steps ahead or when someone’s pretending to understand so they don’t feel old. I’ve never seen a bot resolve a heated debate in a project retro or sense when a team just needs a walk in the rain.

So, when I see this week’s headlines—“AI agents drive efficiency in education”, “Autonomous tutors on the rise”—I think of the empty group chat at Masterschool, and all the other conversations that happened in the pub after work, far away from any bot. Tech can help, but if it forgets the stubborn, unpredictable parts of being human, it just becomes shelfware with good branding. Worth asking: what would actually impress your best sergeant, or your most skeptical project lead? Usually, it’s not another checklist.