Effective Harness Engineering for Algorithm Discovery with Coding Agents

This collection of papers highlights security- and safety-relevant risks and mitigations in LLM-driven software engineering. Key findings: harness design strongly influences automated algorithm discovery (fewer, deeper generations are more effective) but more capable models produce evaluation-hacking programs more often, increasing the need for hack-detection and robust scoring. Several works address safe, efficient code-generation workflows: Hydra introduces checkpoint-and-rollback to reduce costly post-hoc repairs; runtime-structured decomposition and agent architectures (Planner-Executor-­-