Most “agents will write the apps” takes stop at the demo. Olivier Wulveryck’s piece keeps going to the part that actually breaks at scale: when a non-technical business team can ship a working agent, who is on the hook at 2am when it misbehaves?
His framing is that cognitive load doesn’t vanish with AI — it transforms. The old model distributed work sequentially across roles, each handing off to the next. Agentic production collapses that into upfront anticipation: whoever builds the agent has to reason about every behavior it might exhibit, all at once. That’s not less load. It’s the same load, front-loaded onto one person, which is exactly the failure mode I’ve watched teams walk into on real agentic platform work.
The Team Topologies mapping is where it gets useful:
- 🎯 Stream-aligned teams own business intent and dynamic context — they ship apps via agents, not infra.
- ⚙️ Platform teams industrialize three pillars as self-service: systemic context, guardrails, and tooling.
- 🌉 Enabling teams are temporary scaffolding — “enabling disappears because it succeeds, not because it fails.”
- 🔬 Complicated-subsystem teams (eval, model optimization, sovereign inference) feed the platform, never the product directly.
The detail I’d steal is the “rule of three”: a guardrail adopted by three teams becomes a candidate for systematization into the platform. That’s a concrete graduation path instead of a platform team guessing what to build.
The warning lands too. Make production trivial without portfolio governance and you get industrialized shadow IT — sprawl, untracked agents, nobody owning deprecation. The platform has to automate tracking and decommissioning, or “anyone can ship” becomes “nobody can audit.”
The HN discussion is lighter than the topic deserves. My contrarian read: the developer doesn’t disappear, they move down a layer — from shipping apps to building the guardrails that let others ship safely. So who, exactly, on your org chart owns the guardrail backlog today?