New research shows AI coding agents exhibit consistent biases in problem-solving approaches that persist within model families but change across versions, creating novel challenges for production systems.
JSSE passes 99.81% of test262 with zero human code. That's the easy part. Maintainability, harness trust, and the missing layers above conformance are where agent-generated code gets hard.
New research shows LLM monitors systematically go easier on risky actions when those actions are framed as the assistant's own prior output. If you're building coding agents, monitoring is a protocol problem, not a prompt problem.
A coding agent built a chess engine in pure TeX by inventing a virtual machine. The practitioner lesson isn't the stunt. It's what happens when agents redesign your substrate.
