Skip to content

Slow the F Down

armin-ronacher's 2026 blog post + the title-bearing theme of the Pragmatic Engineer Pi episode. A counter-posture to the dominant 2026 "agent maximalism" narrative — the one where every company claims 100% of code is agent-written.

The core math

"Your agent can now spit out 10× more code a day than you can, but it also means it spits out 10× more bugs and errors. Even if it has half your error rate — great, it's not 10× more, it's 5× more. It is still more than you would spit out. So the rate of deterioration in your codebase has now increased. Now go dark factory — take 100 agents doing this to your codebase. What's the end result?"

The ratio that matters is not productivity — it's deterioration rate vs. review throughput. Humans cap at ~1.5k LOC/day review throughput; agents don't. The review side of the equation doesn't scale with the production side.

The "we feel it in our bones" observation

"All the companies claiming that all of their code is now written by agents — yes, we know the quality is garbage. We feel it in our bones when we use your product. It's garbage."

User-perceived product quality is Armin's empirical check on the maximalism narrative. If the output is so great, why do the apps feel worse?

Why it's not anti-agent

Armin ships Pi. Mario ships Pi. Both are arguing against output-maximalism, not against agents. The fix isn't to stop using them — it's to reintroduce the human judgment and review primitives that output-maximalism quietly removed. Pairs with do-not-outsource-thinking (Horthy), judgment-as-the-new-moat (this ingest), and engineer-as-director-of-agents (Zakariasson) as the 2026 "human-in-the-loop is not optional" trio.

Connects to