Simon Willison: November model releases were the inflection point for coding — and now the hard work begins

The inflection moment

It has been reported that Simon Willison told Lenny Rachitsky that the November release of advanced models (referenced as GPT‑5.x alongside Opus 4.5) was the moment AI coding agents went from “mostly works” to “actually works.” Short sentence: that changes everything. Who writes the code when the tooling writes most of it? It’s a seismic shift for how software is built, and Willison—co‑creator of Django and a prolific builder—says we’ve passed the tipping point.

Agents, phone‑first coding, and burnout

Willison explained that he now writes roughly 95% of his code from his phone and that managing agentic systems has left him mentally exhausted by 11 a.m.; it has been reported that this fatigue comes less from coding and more from orchestrating agents. Mid‑career engineers, not juniors, are allegedly the most at risk as day‑to‑day craft work gets automated while higher‑stakes integration and oversight remain human problems. His pragmatic patterns—red/green TDD with agents, templates, and “hoarding” useful snippets—are how he’s staying afloat.

Dark factories and security worries

It has been reported that Willison warned about “dark factories”: systems where AI writes, tests, and ships code with minimal human review. That sounds efficient — until something goes very wrong. He flagged prompt injection as an unresolved security hole and described a “lethal trifecta” of conditions that could produce catastrophic failures; these are serious, systemic risks that deserve more than hand‑waving. And yes, allegedly the “pelican riding a bicycle” has become a tongue‑in‑cheek benchmark for model quality — a reminder that sometimes our tests are almost as weird as the outputs.

What to watch next

The near term, Willison argues, is about pattern adoption and governance: new engineering practices, stronger security thinking, and organizational changes for who owns code quality. Companies and engineers will need to choose whether to lean into agentic workflows or double down on human review. Are we ready for dark factories? The answer will shape the next wave of software jobs—and maybe the next public tech disaster, if we’re not careful.

Sources: lennysnewsletter.com