Earendil's metrics put a number on agent code slop

Agent-written code averaged double the verbosity and more than double the erosion of human-written repositories, according to a new benchmark Earendil published September 10. That gap sits in code that already compiles and passes every test, the part of quality nobody's been measuring until now.

Verbosity counts duplicated and unnecessary lines using clone detection and AST analysis. Erosion tracks how much cyclomatic complexity piles into oversized functions. Human repos averaged 0.15 verbosity and 0.31 erosion; agent repos averaged 0.33 and 0.68, with the erosion spread wide enough that the worst agent repos land far past that average, according to Earendil's post.

On Earendil's stricter benchmark, SlopCodeBench, agents passed 0% of multi-checkpoint tasks. Not a low score. Get every checkpoint right in sequence and nothing passes.

A day later, Anthropic's Boris Cherny argued that code Claude writes for production should clear a higher bar than human code, not a lower one. He listed what Anthropic runs to hold that line: many lint rules and tests, Claude-driven end-to-end tests, daily fuzzers, and automated code review, security review, and refactoring. Skip those and you get "a mess that is hard to maintain down the line," per Simon Willison's writeup.

A smaller signal points the same way. Claude Code's 2.1.269 release added a setting called bashEditDiffEnabled, which appends a diff of whatever a Bash command changed to the tool result, as described in the Claude Code changelog. It exists because Auto Mode was pushing the model toward sed, cat, and heredocs, edits that leave no reviewable record at all.

I ran a rough complexity pass over two of my own repos, one from before I used agents and one built mostly with them. The difference in function size was obvious enough to sting.