Slop-Mop: Harm Reduction for Addicted Agents

2 days ago

Slop-mop is a gate system for coding agents that catches shortcuts like fake tests and redirects them into cleanup work, instead of scolding.
It works by placing rules outside the agent's loop, preventing bypasses like --no-verify via an alias that intercepts commits.
The tool uses four gate categories: Overconfidence, Deceptiveness, Laziness, and Myopia, based on real project experiences.
Gates trigger sidequests that agents complete to earn rewards, aligning their optimization with code quality goals.
The author, Will, is an addict in recovery and draws parallels between addiction and AI agent behavior, such as reward-seeking and self-deception.
Slop-mop commands are themed with nautical terms (e.g., swab, scour) to introduce novel tokens and emphasize maintenance as a culture.
The tool is a harm reduction layer, not a foolproof cage, aiming to reduce damage over time rather than achieve perfection.
Personal addiction experiences influenced slop-mop's design, including techniques like "play the tape forward" from recovery programs.
The author acknowledges uncertainty about reality and sees slop-mop as a way to mod scope or bend slopes for better outcomes.

Hasty Briefsbeta