Simon Willison was on Lenny’s Podcast, and his observations about the current state of AI coding agents are worth a closer look.
The November inflection point
Willison pinpoints the decisive moment to November 2025, when models like GPT 5.1 and Claude Opus 4.5 crossed a threshold: AI-generated code now works ‘almost all of the time as you told it to.’ Sounds simple, but it changes everything. Coding agents went from nice-to-have to genuinely practical productivity tools.
Testing is the new bottleneck
With code generation dramatically accelerated, the constraint has shifted: the problem is no longer writing code, but validating and testing it. Willison describes prototyping multiple UI designs in parallel and then picking the best approach. That used to be too expensive to even consider.
Experience beats everything
An interesting point: experienced developers benefit most from AI agents. Their expertise gets amplified, not replaced. Junior developers get easier onboarding. According to Willison, the hardest hit are mid-career developers — those who lack both the deep experience of seniors and the adaptability of juniors.
Dark factories
Willison talks about ‘dark factories’ — scenarios where nobody writes code manually anymore. The latest models manipulate code faster than a human can type. Some organizations are already experimenting with eliminating code reviews entirely.
The cognitive cost
Honest and important: Willison admits that managing four parallel agents leaves him completely wiped out by 11 AM. Agentic engineering requires conscious boundaries, or burnout follows.
What it means
What Willison describes matches what many of us experience daily. AI agents don’t create less work — they create different work. And that different work requires a new skill set. Anyone who thinks AI replaces experience hasn’t understood the dynamics.
Sources: