Thoughts on slowing the fuck down

My notes

Summary

Mario Zechner (creator of libGDX) argues that the rush to delegate everything to coding agents is producing brittle, unmaintainable codebases at unprecedented speed. The core thesis: agents compound small errors (“booboos”) at a rate humans never could because they have no learning ability, no bottleneck, and only a local view of the system. The fix isn’t to stop using agents - it’s to slow down, stay in the code, write architecture by hand, limit daily agent output to what you can actually review, and treat agents as tools for boring/exploratory work rather than autonomous builders.

Key Insight

• Booboos compound without a bottleneck. A human makes errors slowly and feels the pain, so they fix things before rot sets in. Agents produce errors at machine speed with no pain signal. By the time you notice, the codebase is unrecoverable. This is the central mechanism behind “agentic technical debt.”
• Agents are merchants of learned complexity. Their training data is full of cargo-cult enterprise patterns. Left unsupervised, they reproduce the same bloated abstractions that take human teams years to create - but in weeks. Two humans with an agent army can reach enterprise-grade mess faster than a 50-person team.
• Agentic search has low recall at scale. As the codebase grows, agents increasingly fail to find all relevant code before making changes. This causes duplication, inconsistency, and architectural drift - the exact problems that make the codebase harder to search, creating a death spiral.
• Tests become untrustworthy too. If agents wrote the code, they also wrote the tests. When the codebase is compromised, so is the test suite. The only reliable check becomes manual testing - the opposite of what automation promised.
• Good agent tasks share specific properties. They can be scoped without full system understanding, the agent can evaluate its own work (closed loop), and the output isn’t mission-critical. Everything else needs human involvement.
• Architecture, API design, and system gestalt must be hand-written. The friction of writing these yourself is a feature, not a bug - it forces you to understand what you’re building and why. This is where experience and taste matter, and current models can’t replace either.
• Set daily agent output limits. Cap the amount of generated code per day to what you can genuinely review. Speed of generation is meaningless if review becomes the bottleneck you ignore.