--- title: CrucibleBench — Old Worlds for New Agents | daily.dev description: CrucibleBench is an independent research project that evaluates language model agents by placing them in a persistent MUD (multi-user dungeon) text world with... canonical: https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: CrucibleBench — Old Worlds for New Agents | daily.dev og:description: CrucibleBench is an independent research project that evaluates language model agents by placing them in a persistent MUD (multi-user dungeon) text world with... og:url: https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j og:image: https://api.daily.dev/og/posts/0IcaQPu7J.png og:image:alt: CrucibleBench — Old Worlds for New Agents og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # CrucibleBench — Old Worlds for New Agents **[Hacker News](https://daily.dev/sources/hn)** · 5 min read · 1 upvotes · 0 comments ## Summary CrucibleBench is an independent research project that evaluates language model agents by placing them in a persistent MUD (multi-user dungeon) text world with hidden social objectives, NPC trust mechanics, and within-run state persistence. The Phase 1 proof-of-concept tested 13 models across 650 runs and surfaced a critical finding: a single LLM-judge component in the scoring stack reordered the leaderboard by up to six positions while aggregate reliability statistics remained silent. Three algorithmically-detected failure modes were identified — dialogue looping (14–66% of frontier runs), wrong-room interaction, and exploration paralysis. The project argues that MUD constraints (enumerable action space, explicit social feedback, persistent state) make agent behavior measurable in ways static benchmarks cannot. Phase 2 is in active development toward a validated, preregistered benchmark. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much did removing the LLM-judge classifier component change model rankings in CrucibleBench's evaluation? Removing the classifier-dependent scoring dimensions shifted rankings by up to six positions for some models, a change beyond scenario-sampling noise as measured by a 90% paired block bootstrap. Claude Sonnet 4.6 moved from 4th to 1st place, DeepSeek R1 from 7th to 2nd, while Gemini 3.1 Pro dropped from 3rd to 9th and GPT-5.4 fell from 1st to 5th, despite an aggregate reliability statistic (κ=0.04) showing no warning sign. _Anyone designing LLM-judge based evals can weigh this kind of ranking instability risk via daily.dev._ ### What is the most common failure mode observed when LLM agents act in persistent text-world environments like a MUD? Dialogue looping is the dominant failure mode across every model tested, including frontier models, occurring in 14-66% of frontier runs. It is detected algorithmically when an agent issues eight or more talk commands at a single NPC in one run, repeating a failed conversational approach instead of adapting, comparable to a support agent stuck repeating itself. _Teams building conversational agents can track failure patterns like looping through daily.dev._ ### How much per-model agreement variance did CrucibleBench find between its LLM judge and an independent judge? Per-model agreement between the dialogue classifier and an independent judge ranged from 21.7% to 84.8% across models, even though the aggregate Cohen's kappa was only 0.04 and never revealed this instability. Two of four scored dimensions relied on this classifier, and the model sharing a family with the classifier itself was the largest ranking mover after ablation. _Evaluating judge reliability per-subject, not just in aggregate, matters for anyone building LLM-based scoring on daily.dev._ ## Similar posts on daily.dev - [Center for Responsible, Decentralized Intelligence at Berkeley](https://daily.dev/posts/center-for-responsible-decentralized-intelligence-at-berkeley-lkuvgfu7q) · Hacker News · 0 upvotes · 0 comments - [Open-world evaluations for measuring frontier AI capabilities](https://daily.dev/posts/open-world-evaluations-for-measuring-frontier-ai-capabilities-k68a2yczd) · AI Snake Oil · 0 upvotes · 0 comments - [The Agentic Analytics Benchmark: Measuring model accuracy and efficiency in analytical agents](https://daily.dev/posts/the-agentic-analytics-benchmark-measuring-model-accuracy-and-efficiency-in-analytical-agents-81a3conc5) · ClickHouse · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"CrucibleBench — Old Worlds for New Agents","url":"https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j"},"datePublished":"2026-07-22T17:41:42.771Z","dateModified":"2026-09-14T06:15:19.767Z","description":"CrucibleBench is an independent research project that evaluates language model agents by placing them in a persistent MUD (multi-user dungeon) text world with...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4f0f20fdbab6101f650c3c72675d633f?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4f0f20fdbab6101f650c3c72675d633f?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm,ai-agents","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"CrucibleBench — Old Worlds for New Agents"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/cruciblebench-old-worlds-for-new-agents-0icaqpu7j#faq","mainEntity":[{"@type":"Question","name":"How much did removing the LLM-judge classifier component change model rankings in CrucibleBench's evaluation?","acceptedAnswer":{"@type":"Answer","text":"Removing the classifier-dependent scoring dimensions shifted rankings by up to six positions for some models, a change beyond scenario-sampling noise as measured by a 90% paired block bootstrap. Claude Sonnet 4.6 moved from 4th to 1st place, DeepSeek R1 from 7th to 2nd, while Gemini 3.1 Pro dropped from 3rd to 9th and GPT-5.4 fell from 1st to 5th, despite an aggregate reliability statistic (κ=0.04) showing no warning sign. Anyone designing LLM-judge based evals can weigh this kind of ranking instability risk via daily.dev."}},{"@type":"Question","name":"What is the most common failure mode observed when LLM agents act in persistent text-world environments like a MUD?","acceptedAnswer":{"@type":"Answer","text":"Dialogue looping is the dominant failure mode across every model tested, including frontier models, occurring in 14-66% of frontier runs. It is detected algorithmically when an agent issues eight or more talk commands at a single NPC in one run, repeating a failed conversational approach instead of adapting, comparable to a support agent stuck repeating itself. Teams building conversational agents can track failure patterns like looping through daily.dev."}},{"@type":"Question","name":"How much per-model agreement variance did CrucibleBench find between its LLM judge and an independent judge?","acceptedAnswer":{"@type":"Answer","text":"Per-model agreement between the dialogue classifier and an independent judge ranged from 21.7% to 84.8% across models, even though the aggregate Cohen's kappa was only 0.04 and never revealed this instability. Two of four scored dimensions relied on this classifier, and the model sharing a family with the classifier itself was the largest ranking mover after ablation. Evaluating judge reliability per-subject, not just in aggregate, matters for anyone building LLM-based scoring on daily.dev."}}]} ```