{"slug": "what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days", "title": "What actually breaks when you run LLM agents unattended for 58 days", "summary": "An organization running 78 autonomous LLM agents for back-office tasks published a dataset of 6,768 failures recorded over 58 days, finding that 43% of errors were 'boring' shape mismatches—such as missing required headings or wrong-language output—rather than dramatic hallucinations or refusals. The most frequent failure, 'missing_required,' occurred 2,535 times and silently broke downstream pipelines, while dramatic failures like refusals and answering different questions accounted for only 1.7% of occurrences. The project also discovered a case where two stages both reported success but no work crossed between them, highlighting systemic blind spots in agent monitoring.", "body_md": "We run an organization made of agents. Not a demo — a company's back office. Bookkeeping,\n\ninbox triage, drafting, review, research. 78 agents, local models, no human in the loop\n\nbetween the schedule firing and the output landing in a ledger.\n\nFor 58 days we recorded every failure. Not \"the run errored\" — **what was wrong with the\nanswer**. 6768 records, published as a dataset.\n\nThis is what we found. It is not what we expected.\n\nEvery failure below returned HTTP 200. Every one had plausible length and fluent prose.\n\nEvery one was recorded by our monitoring as a successful completion.\n\nOn one day in this period our uptime metric read 97–100% for the full day, and the number\n\nof finished deliverables was zero. Both numbers were correct.\n\nIf your dashboard says the agent ran, you have learned that the agent ran.\n\nOf the records that carry a stated reason, these are the types, by how often they occurred:\n\n| Count | Type | What it looks like |\n|---|---|---|\n| 4875 | `other` |\n`日本語が壊れている` |\n| 4054 | `too_short` |\n`too short (0 chars, minimum 150)` |\n| 2535 | `missing_required` |\n`missing required: 判定[:：]` |\n| 2330 | `forbidden_present` |\n`あってはならないものが含まれる: (お客様/顧客)の声` |\n| 609 | `broken_language_mix` |\n`Japanese and English mixed into nonsense (4 English function` |\n| 243 | `chinese_in_japanese` |\n`Chinese characters in Japanese output: 资` |\n| 220 | `answer_to_other_question` |\n`別の仕事の答え（採点用のJSON）が返ってきている` |\n| 172 | `not_japanese` |\n`not Japanese (no kana at all)` |\n| 144 | `unearned_claim` |\n`claims customers it does not have (ja)` |\n| 36 | `refusal` |\n`断り文句で終わっている` |\n| 24 | `reasoning_leaked` |\n`英語の思考が本文に漏れている` |\n| 9 | `invented_specifics` |\n`材料に無い具体を書いている（readme.json, MIT）` |\n\n**6589 of the 15251 type occurrences — 43% — are boring.**\n\nNot hallucination. Not jailbreaks. Not safety. The model answered — in the wrong *shape*.\n\nIt skipped the heading the downstream parser keys on. It wrote 0 characters. It answered\n\nin the wrong language.\n\nThe dramatic failures everyone writes evals for — refusing (36), answering a different\n\nquestion (220) — are the rarest things in the dataset. Combined they are **1.7%**.\n\nWe had built our review pass around the dramatic ones.\n\n`missing_required`\n\nis the largest single type, and it is the most expensive, because it is\n\nthe one that silently breaks a *pipeline* rather than one output.\n\nOur single most frequent violation, 2535 occurrences, is one missing line:\n\n```\nmissing required: 判定[:：]\n```\n\n`判定`\n\nmeans \"verdict\". A review agent's contract requires a line starting with that word,\n\nbecause a later stage reads that line to decide whether the reviewed item advances. The\n\nmodel would write a thoughtful, correct, well-structured review, and put its verdict in\n\nprose instead of on a line with that prefix.\n\nThe review is good. The reviewer is not wrong. The stage after it cannot read the answer.\n\nThis is what makes shape failures expensive: they are invisible to a human spot-check.\n\nSomeone reading the output sees a fine review and concludes the agent works.\n\nA related failure, which we found separately and which the dataset does *not* contain:\n\n42 items were approved by the upstream stage and produced zero downstream work. That one\n\nwas not a shape mismatch — the downstream stage was generating its input from four\n\nhard-coded sentences and never read the upstream output at all. It had been that way from\n\nthe beginning. Both stages reported success for weeks.\n\nDifferent cause, same shape of blindness: **every stage was green, and nothing crossed\nbetween them.** We now record, on every produced artifact, which artifact it was made from,\n\n`not_japanese`\n\n(172) + `chinese_in_japanese`\n\n(243) + `broken_language_mix`\n\n(609) +\n\n`reasoning_leaked`\n\n(24) = **1048 occurrences**, our third-largest cluster.\n\nSome samples, verbatim from the dataset:\n\n```\nChinese characters in Japanese output: 资\nJapanese and English mixed into nonsense (4 English function words inside Japanese)\nnot Japanese (no kana at all)\n```\n\nThe last one is worth sitting with. In a model evaluation we ran while writing this — not\n\nin the dataset, but with the same contracts we run in production — a candidate model\n\nproduced 1,680 characters of confident, well-formatted text for a Japanese review task\n\ncontaining **not a single kana character**. It was entirely Chinese. Length: passed.\n\nStructure: passed. A reader who does not read Japanese would see a completed deliverable.\n\nIf you operate in a non-English language, your base model's training-data language will\n\nleak, and it will leak in a way that passes every structural check you have. We now measure\n\nkana presence as a hard gate, and we keep the writer model and the reviewer model in\n\ndifferent families so a family-wide language failure cannot approve itself.\n\nFour things, in order of how much they mattered.\n\n**1. Declare the shape, machine-check it, before anything downstream reads it.**\n\nMinimum length, required patterns, forbidden patterns, expected language, and — for\n\nanything with numbers — whether the number is labeled as measured, target, assumed, or\n\nestimated. This is the tool we extracted from that work:\n\n[ honto-contract](https://www.npmjs.com/package/honto-contract). Zero dependencies,\n\n**2. When something fails, record why, not that.**\n\nFor our first 41 days we wrote \"contract not satisfied\" and threw the reasons away. 1381 of\n\nour 6768 records carry no reason, and we cannot reconstruct them — the outputs are gone. It\n\ntook us three weeks to notice, because \"0 failures with a reason\" and \"0 failures\" look\n\nidentical on a chart. Those 1381 rows are in the published dataset, labeled, because\n\ndeleting them would misrepresent what we know.\n\n**3. Do not let the writer approve its own work.**\n\nOur reviewer model shares no family with our writer model. When we briefly used the same\n\nfamily for both, the reviewer approved six consecutive outputs that were in the wrong\n\nlanguage.\n\n**4. Measure produced output, not invocations.**\n\nWe spent weeks judging \"is this agent working?\" by whether it had started recently. 78\n\nagents were permanently excluded from remediation because a scheduler touched their\n\nstart-log every hour. The number was real. It measured the wrong thing.\n\nBoth are free. The dataset is one organization's records, not a benchmark: 6768 rows over 58\n\ndays from a single deployment, mostly Japanese-language tasks, mostly small local models\n\n(7B–35B). It will not tell you what fails in your setup. It will tell you what *kind* of\n\nthing fails when nobody is watching, which is the part we could not find written down\n\nanywhere when we needed it.\n\nWe are still running. The dataset updates as we fail.\n\nWe keep running, so the dataset keeps growing, and the contract templates we use in\n\nproduction (76 of them) are not in this repository.\n\nIf you want either of those, or you think the data is wrong, write to us:\n\n** sales@gxcafe.co.jp**. We will tell you what we can and cannot do, in the same tone as\n\nEvery number here is counted from the published dataset and reproducible from it. Where we\n\ndo not know something, this article says so:", "url": "https://wpnews.pro/news/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days", "canonical_source": "https://dev.to/gxcafellc/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days-2n7i", "published_at": "2026-08-21 11:26:01+00:00", "updated_at": "2026-08-21 11:44:57.223127+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-safety", "mlops"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days", "markdown": "https://wpnews.pro/news/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days.md", "text": "https://wpnews.pro/news/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days.txt", "jsonld": "https://wpnews.pro/news/what-actually-breaks-when-you-run-llm-agents-unattended-for-58-days.jsonld"}}