Agent Slop¶
Summary¶
The accumulation of low-quality, overly complex, and poorly understood AI-generated code that results from agents working at high velocity without human friction. Term coined by Mario Zechner as "boooos" — compound errors from serial learning with no bottlenecks and delayed pain.
The Problem¶
Agents compound errors because: 1. No bottlenecks — 10 agents can add far more "boooos" per day than a single human 2. Delayed pain — humans feel pain and refactor; agents happily dig deeper 3. Serial learning — each agent decision is local, especially when codebase exceeds context window 4. Learned complexity — agents learn from the internet, where 90% of code is "our old garbage"
How Agents Create Enterprise-Grade Complexity¶
In 2 weeks with 2 humans and 10 agents: - Agents add abstractions everywhere that intertwine - Duplication accumulates - Backwards compatibility layers multiply - Defense in depth — layers of unnecessary safety code - Result: enterprise-grade complexity that no human understands
The Codebase Death Spiral¶
- Agents generate code humans can't review at scale
- Codebase becomes too complex for agents to understand
- Agents patch locally, break globally
- Tests are written by agents — can't trust them either
- Users scream about bugs
- Agents are now overwhelmed by the codebase they created
- You haven't read the code — you can't fix it. Agents can't fix it either. Game over.
Mario Zechner's Advice¶
- Non-critical code: wipe slop ahead
- Critical code: read every line
- Important things: write by hand — the friction builds understanding
- Say no — your most valuable capability
- Fewer features, but the ones that matter
- Slow the fuck down
Relationship to AI Slop¶
Ai Slop refers to AI-generated content broadly. Agent slop is specifically about code — the compound effect of agents writing, reviewing, and testing their own code without human understanding of the system.