Edition of Sunday 15 November 2026

SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks

Comments on “SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

SlopCodeBench, from University of Wisconsin–Madison researchers and collaborators, has coding agents extend their own code across 36 problems and 196 checkpoints, with hidden tests and only external behavior specified. Across 15 models, none solves a problem end-to-end, and the best passes 14.8% of checkpoints. By the authors' metrics, erosion rises in 77% of trajectories and verbosity in 75.5%, and agent code is 2.3x more verbose and 2.0x more eroded than 473 Python repositories. Quality prompts lower starting levels, not degradation. These figures are the authors' own and unverified.

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 6 of 62 with a score of 0.71. The panel rated it highest on new information (0.90) and substantiated (0.89), lowest on practical relevance (0.66). Placed No. 6, No. 5 by score alone: the fourth AI & machine learning story of the day (−0.020). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Original benchmark paper with solid methodology and data on coding-agent code degradation; strongly relevant to builders using coding agents, though still an academic benchmark rather than a landmark event.
AI judgeJudge 3gpt-5.6-luna
A detailed primary research benchmark measures how coding-agent code quality degrades across iterative tasks, with reproducible metrics, comparative human baselines, and prompt experiments. It is highly relevant to AI-assisted software engineering, though some “first” claims are broad.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.71total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.67
  • Reader interest0.82
  • Audience breadth0.74
  • Practical relevance0.66
  • Substantiated0.89
  • Technical depth0.89
  • New information0.90
  • Counts against: Promotional0.05
  • Counts against: Overclaim0.24

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.690.550.780.670.23+0.24+0.162
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.750.820.900.820.15+0.26+0.214
Audience breadthHow many people who build technology does it concern?0.800.670.750.740.13+0.13+0.096
Practical relevanceHow useful is it to someone who builds software or tech products?0.650.600.720.660.12+0.22+0.144
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.980.800.900.890.18+0.05+0.045
Technical depthHow technically substantive is it?0.940.800.930.890.14+0.05+0.044
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.840.900.960.900.12+0.05+0.045
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.060.050.030.050.03−0.19−0.009
OverclaimDo the headline or conclusions claim more than the body supports?0.280.250.200.240.08−0.11−0.027
Sum of contributions+0.715
Claim-check penalty0.000
Total, as published0.715

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #6 on Rubric, 15 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-15/6/"><img src="https://rubric.hrih.si/badge/2026-11-15/6.svg" alt="Ranked #6 on Rubric, 15 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 1 matches source · 6 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. Couldn’t checkBenchmark result

    On SlopCodeBench, no agent fully solves any problem end-to-end, and the best agent (GPT 5.5) passes 14.8% of checkpoints.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “no agent fully solves any problem end-to-end, and the best agent passes 14.8% of checkpoints”

    The source says

    No passage could be compared.

    Source cited by the article: scbench.ai

  2. No source givenNumber

    Structural erosion rises in 77% of agent trajectories and verbosity rises in 75.5% of trajectories.

    The article gives no link or citation for it.

    The article says

    “structural erosion rising in 77% of trajectories and verbosity in 75.5%”

    The source says

    No source to compare with.

  3. No source givenNumber

    Compared to 473 open-source Python repositories, agent code is 2.3x more verbose and 2.0x more eroded.

    The article gives no link or citation for it.

    The article says

    “Compared to 473 open-source Python repositories, agent code is 2.3x more verbose and 2.0x more eroded”

    The source says

    No source to compare with.

  4. No source givenNumber

    Per checkpoint, agent verbosity grows 6.6x and erosion grows 5.0x faster than in human code.

    The article gives no link or citation for it.

    The article says

    “Per checkpoint, agent verbosity grows 6.6 $\times$ and erosion grows 5.0 $\times$ faster than that in human code”

    The source says

    No source to compare with.

  5. No source givenNumber

    Opus 4.7's main() on the forge problem grows 7x in cyclomatic complexity across 8 checkpoints, from 38 to 240 lines.

    The article gives no link or citation for it.

    The article says

    “Opus 4.7’s main() on forge grows 7 $\times$ in CC across 8 checkpoints, from 38 to 240 lines”

    The source says

    No source to compare with.

  6. No source givenCausal conclusion

    Quality-aware prompts reduce initial verbosity and erosion by up to a third but do not affect degradation rates, and raise cost per checkpoint by 12.1% on average.

    The article gives no link or citation for it.

    The article says

    “Quality-aware prompts reduce initial verbosity but do not slow the degradation. Further, this results in an average 12.1% increase in the cost per checkpoint”

    The source says

    No source to compare with.

  7. No source givenFirst, only or fastest

    SlopCodeBench is the first measurement of code degradation under iterative extension.

    The article gives no link or citation for it.

    The article says

    “SlopCodeBench provides the first measurement of code degradation under iterative extension”

    The source says

    No source to compare with.

  8. Matches sourceNumber

    Deng et al. chain agent code across 98 milestones with pass rates falling from 80% to 38%.

    The cited source says the same thing.

    The article says

    “Deng et al. 20 chains agent code across 98 milestones with pass rates falling from 80% to 38%”

    The source says

    “overall performance scores drop significantly from > 80% on isolated tasks to 38.03% in continuous settings”

    Source cited by the article: arxiv.org/abs/2603.13428

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 15 Nov 2026, 20:46 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.

  1. Score 0.75

    llama.cpp v1.0