Someone put a number on agent code slop: twice as verbose and twice as eroded as human repos
earendil.com (266 points on HN)·high signal
Sebastian at Earendil published two measurable metrics on 10 September, verbosity (duplicated and unnecessary lines via AST-Grep and clone detection) and erosion (complexity concentrated in large functions via cyclomatic complexity). Human repos averaged 0.15 ± 0.06 verbosity and 0.31 ± 0.17 erosion; agent-written code hit 0.33 ± 0.10 and 0.68 ± 0.20. His argument is that LLMs now produce 'almost perfectly correct code' while failing on everything past correctness, backed by SlopCodeBench showing 0% pass rates on strict multi-checkpoint evaluation.