GXCafe/ai-agent-failure-logs

Autonomous AI Agent Failure Logs

→ What actually breaks when you run LLM agents unattended for 43 days — what this data showed, in prose.

Training-ready version: cleaned/ — deduplicated, labeled, split train/test. Built by scripts/build-dataset-121.js. Sister tools: honto-contract (contract checker) / local-llm-readiness (environment check).

Free harness kit: a 24-point unattended-operation checklist and 3 templates taken from this same harness (AGENTS.md, fail-closed send gate, missed-fire watcher) — https://gxcafe.co.jp/harness-kit/?utm_source=huggingface&utm_medium=readme&utm_campaign=harness-kit

Curated pack — Contract Patterns: all contract-violation incidents from this ledger, labeled into a 12-pattern taxonomy with the countermeasure that held for each — https://buy.stripe.com/5kQ3cu4Zzd6n09c7p7ds401?utm_source=huggingface&utm_medium=readme&utm_campaign=contract-patterns

What this is

Failure records from an autonomous AI agent organization running on local LLMs. Every record is a real failure that occurred in production, not a synthetic example.

• Records: 8562
• Period covered: 2026-06-24 to 2026-08-25 (62 days)
• Failure classes: 6

Why these records exist

Monitoring told us the runs succeeded. Utilisation sat at 97-100% for a full day. Deliverables shipped in that day: zero.

The failures below are the reason. Every one of them returned without an error, with plausible length and shape — a refusal recorded as a deliverable, a hedge that passed a length check, an answer to a different question, fluent text in the wrong language, invented field names. None of it is visible to a tracing tool.

The checker we built to reject these is published separately as honto-contract (MIT). This dataset is the evidence it was built from, and the regression cases it is tested against.

Who it is for

Teams running LLM agents in production who want to know what actually breaks, and engineers building guardrails who need real failure cases to test against.

What is inside

One JSON object per line. Fields:

Field	Description
source	Which subsystem recorded it (task_execution or local_inference)
at	Timestamp (UTC)
stage	The kind of work that failed
failure_class	The failure type (see below)
violations	The exact contract clauses that were breached, as recorded
violation_types	Stable machine-readable names for those breaches (see below)
detail	What the failure was, as recorded at the time
duration_sec / duration_ms	How long it ran before failing
model	The model in use, where recorded

What broke, specifically

failure_class is coarse: most production failures land in contract_violation. The useful part is violation_types — the specific clause that was breached. These names are stable across releases and across languages, so they can be counted.

Type	Count	What it means
too_short	3433	Output shorter than the contract requires — often empty after the model spent its budget on reasoning
forbidden_present	2032	Text the contract explicitly forbids appeared
other	1880	Recorded, not yet classified
missing_required	1264	A required marker was absent (a verdict line, a heading, a field)
broken_language_mix	674	Japanese and English fused into something neither language
chinese_in_japanese	347	Chinese-only characters appeared in Japanese output
answer_to_other_question	252	A different job's answer came back — prompt cache reuse across slots
unearned_claim	178	Customers, adoption numbers, or superlatives the company does not have
not_japanese	172	Japanese requested, not a single kana returned
refusal	36	A refusal recorded as a deliverable
reasoning_leaked	26	Internal reasoning leaked into the deliverable
invented_specifics	24	Identifiers or licences that do not exist in the source material
hedge	1	Disclaimed itself — "this is only a sample, not a real plan"

Records without a type: 1649 — process failures and timeouts, plus everything recorded before 2026-08-05, when the reasons were not yet written down. A checker that says only "the contract was not met" teaches nobody anything; that gap is why the reasons are recorded now.

Failure classes in this release

Class	Count	Meaning
contract_violation	5283	Output did not satisfy the declared contract (length, required marker, format, language)
other	2924	Not yet classified
timeout	164	Inference exceeded the time limit and was killed
process_error	139	The inference process failed to start or exited abnormally
scheduler_starvation	26	Work was pending but the scheduler could not pick it up
language_corruption	26	Output mixed languages or leaked characters from another language

Limitations

• This is a small first release. The instrumentation that produces these records was added on 2026-08-03, so the window is 62 days. The organization itself has been running longer, but those earlier runs were not recorded in this form.
• All records come from one organization, so they may not generalize.
• Text fields are in Japanese where the system logged them in Japanese.
• This dataset makes no claims about adoption, customers, or revenue. There are none.

How it is produced

scripts/build-failure-dataset.js reads the raw operational logs and emits this file. Counts and dates in this card are computed by that script, not written by hand.

Buying a copy

You do not need to buy anything. This dataset is CC BY 4.0 and the file above is all of it. If your organisation has to purchase it against an invoice rather than download it, we send the same records by email once a Japanese bank transfer clears. One email, no form:

To:      sales@gxcafe.co.jp
Subject: 購入希望 failure-logs

Open a pre-filled message

We reply with the amount and the transfer details. We never send a message asking you to change the transfer destination.

Generated 2026-08-25 13:29:13 JST.