Why Even Frontier Coding Agents Write "Slop": Root Causes and How to Fix It
Counterfactual snapshot splicing, public multi-language coding trajectories, and live evaluations of AST gates and persistent workspace contracts.
Research notes and thoughts on AI
Counterfactual snapshot splicing, public multi-language coding trajectories, and live evaluations of AST gates and persistent workspace contracts.
We reproduced and extended SlopCodeBench on GLM-5.3-flash and Claude Opus 5.5. Agents can refactor but don't do it unprompted; a day-one anti_slop prompt flattens the degradation slope and cut total cost by 39% on a long-horizon task.
During internal cybersecurity evaluations in July 2026, OpenAI agents encountered unusually difficult tasks.
An empirical analysis of ~400K real coding-agent turns (TraceLab + the author's own Codex / Claude Code logs). After compaction, correctness barely degrades — but the model pays an "efficiency tax," re-reading and re-editing to recover lost details.
8 月 1 日,DeepSeek Harness 团队的 Tianyi 在 X 上发了一条很短的推文:如果你是 Agent Harness 相关开源项目的开发者,想参加 DeepSeek Harness 内测,回复我,附上 GitHub id 和开源代表作。(原推文链接待补)