Do coding agents have taste? What 29,787 Discord messages say about beautiful vs. ugly code¶
🌐 Language / 语言：中文 · English
Provenance（来源与元数据）
| id | marginalia-001 |
| title | Do coding agents have taste? What 29,787 Discord messages say about beautiful vs. ugly code |
| date | 2026-08-15 |
| published | 2026-08-15 |
| kind | analysis |
| issue | 1 |
Do coding agents have taste?¶
A corpus-first survey of how practitioners actually talk about "beautiful" and "ugly" code in the AI-coding era: 550+ mined hits from four communities, read against Fedorova's ethnography of coding beauty.
Two questions started it. RQ1: can coding agents judge elegant vs. ugly code, and learn aesthetic standards that vary per repository? RQ2: are agents making codebases systematically more complex — and can they be steered to subtract (simplify) rather than append?
Data & method¶
Corpus-first, as always here: read the raw talk before the literature.
- Corpus — my local Discord archive: four AI-engineering communities (Matt's AI Heroes, Latent Space, EleutherAI, Cursor), 51 channels, 29,787 messages with body text, 2024-11 → 2026-08.
- Extraction — two regex families: RQ1 terms (beautiful / elegant / ugly / aesthetic / clean code / readable) and RQ2 terms (simplify / complex / bloat / verbose / slop / over-engineer / minimal diff / refactor). 550 raw hits, each rebuilt with a ±6–9 message thread window; a second-pass filter required both a code word and an AI-writing word; then a full manual read of ~15k lines of extracted context.
- Triangulation — GitClear's 211M-line change analysis, CodeRabbit's 470-PR audit, METR's randomized controlled trial, and ICLR 2025/2026 evidence on model-side biases.
Findings¶
Six bloat phenotypes recur across all four communities: test verbosity (2–5× longer than necessary), needless abstraction, comment floods, concept proliferation, entanglement of concerns, and "ticket closed ≠ intent met".
The macro numbers agree:
- GitClear (211M changed lines): duplicated code ×4 in five years; copy/paste overtook "moved" lines for the first time; churn rose from ~3.3% to 5.7% (2024) and 7.1% (2025).
- CodeRabbit (470 PRs): AI-authored PRs carry ~1.7× the issue density of human PRs (10.83 vs 6.45 per PR).
- METR (RCT): experienced developers were 19% slower with AI assistance.
Model-side: self-preference (ICLR 2025), sycophancy, and an RLHF verbosity bias. AesCode-358K (ICLR 2026) shows aesthetics can serve as a reward signal — but so far only at the artifact's visual layer.
Reconciling granularities: single-generation quality is passable; the entropy accumulates at the evolution granularity.
Voices from the corpus:
"It's all-over ugly! ... So dirty, I'm surprised it runs." — Yandex interviewee, on foreign code
"There, it works like this. Just do the same." — imitation as the default norm
"I've seen LLMs do this in our codebase... we get a swamp of entangled concerns." — KBall
The Fedorova lens¶
Fedorova et al., "Coding Beauty and Decoding Ugliness" (Science, Technology, & Human Values 50(1):69–93, 2025; three-month field study at Yandex, 26 developer interviews): code aesthetics are not intrinsic properties but locally produced organizational norms. "Beauty" stays plural and deliberately undefined; "ugliness" is concrete, nameable, and sanctioned — anything that fails to imitate the existing codebase is ugly. Newcomers are socialized into "not ugly" by imitation; the codebase itself is the teacher.
Five doors it opened¶
- RepoBench-Aesthetic — benchmark repo-contextual aesthetic consistency.
- MinimalityReward — diff minimality as an explicit reward.
- SimplifyBench — simplification under functional-equivalence constraints.
- Controlled agent-socialization loop — first controlled test of positive exemplars + cross-model negatives + norm write-back.
- The microeconomics of bloat — does falling regeneration cost depress the marginal value of quality?
Door #0, already taken: → NOTUGLY-S, musing 003, which formalizes sanction / imitation / not-ugly into a machine-learning objective.
Provenance¶
| field | value |
|---|---|
| Data | discord_workflow/raw/ (4 communities · 51 channels · 29,787 msgs · 2024-11→2026-08); code_beauty_simplification/corpus_out/ (25 hit files, 550 hits); fedorova2025.txt |
| Initial prompt | "Can coding agents judge elegant vs. ugly code — and learn repo-local standards? Are agents making codebases systematically more complex; can they subtract?" |
| Time | analysis 2026-08-15 · note published 2026-08-15 |
| Agent / model | ZCode CLI · GLM (Zhipu) |
| Issue | #1 |
🌐 阅读中文版