Thinking in Systems, Shipping in Loops

At Rails World in Austin on 23 September 2026, David Heinemeier Hansson made the blunt case that writing code by hand has stopped being economically useful for most programmers at most companies. The quote landed hard because it wasn’t framed as a distant future. It was framed as now, with the end point pushed even further: virtually all domains, virtually all programmers, virtually all companies by year’s end.

That’s the shift in plain terms. Software engineering is moving from typing code to designing the system around the code — the rules, checks, and constraints that decide whether AI’s output is any good. The work is less about crafting every line and more about steering the machine so the lines it produces are safe, correct, and reusable.

The source points to Artemis as the clearest example. Its engineers went from merging 2 pull requests a day in January to 6 by May and 16 by August, adding up to 30,000 in eight months. Dan Shiebler, the company’s CTO, describes a clean split: agents write every line on the platform, while engineers set the constraints, design the systems, and review what comes out.

Lauren Tan on SpaceXAI’s Grok team is doing something just as telling, shipping 2,000 pull requests a month to production. That pace only works because the review loop is built in. She doesn’t credit a model breakthrough; she credits verification, the machinery that lets AI check its own work and keep moving.

Donella Meadows’ old systems ideas fit neatly here. The source leans on three properties — resilience, self-organization, and hierarchy — to explain what good agentic systems need. Resilience means multiple checks, not one brittle gate. Self-organization means the system learns from failure. Hierarchy means skills, tools, and sandboxes stack into reusable parts once they’ve been proven. The punchline is simple: the scarce craft is no longer typing. It’s deciding how intelligence gets steered.