Member-only story
Mario Zechner liked Claude Code. Then he watched it get worse in a way that’s specific to how agent tools tend to fail: not through any single bug, but through accumulation. New features shipped, system prompts changed between releases without notice, tool definitions got modified or removed underneath him, and the tool started inserting what he called system reminders, unexplained context injected mid-session that the model had no way to evaluate as relevant or not. He didn’t have visibility into any of it. He just noticed the tool’s behavior kept drifting.
His response wasn’t a bug report. It was a rewrite from zero, published as a long essay titled “Building pi in a World of Slop”. The core argument: mainstream coding harnesses control your context behind your back, and every feature they bolt on is one more thing that can silently change how the agent behaves between one session and the next. His fix was the opposite of feature-complete. Seven built-in tools, no more. A system prompt under 1,000 tokens. And an agent that, if it needs a capability it doesn’t have, writes that capability itself.
Zechner’s essay names the specific failures he was reacting to, and they’re worth listing exactly because they’re concrete rather than vague dissatisfaction: system prompts that changed on every release…