The failure mode your dashboard cannot see: 43% of agent errors are wrong-shaped output at HTTP 200

An operator running a company's back office on agents has published a 58-day failure log covering 78 agents on local models, with no human between the schedule firing and the output landing in a ledger ^[1]. Every failure in that log returned HTTP 200, arrived with plausible length and fluent prose, and was recorded by monitoring as a successful completion ^[2].

The line that should settle any argument about uptime as a proxy for work: on one day in the period the uptime metric read 97-100% for the full day while the count of finished deliverables was zero, and the author's position is that both numbers were correct ^[3].

The log holds 6768 records over the 58 days ^[4], roughly 117 recorded failures a day ^[1]. Among records carrying a stated reason there are 15251 type occurrences, and 6589 of them, 43%, are shape failures: a skipped heading that the downstream parser keys on, zero characters written, or an answer in the wrong language ^[5]. The failures that safety evals are built around are the rare ones. Refusals number 36 and answering a different question 220, together 1.7% of occurrences ^[6], which is 256 events ^[2] and makes the boring category about 26 times more frequent ^[3].

The largest single type is missing_required, and the author argues it is the most expensive because it silently breaks a pipeline rather than one output ^[7]. The single most frequent violation, 2535 occurrences, is one absent line: a review contract requires a line beginning with the Japanese word for verdict, because a later stage reads that line to decide whether the reviewed item advances, and the model instead wrote a correct, well-structured review with its verdict in prose ^[8]. That one violation is 17% of all type occurrences and 38% of the shape cluster ^[4]. It is also invisible to a human spot-check, because a reader sees a good review and concludes the agent works ^[9].

Worse, and separate from the dataset: 42 items were approved by an upstream stage and produced zero downstream work, because the downstream stage was generating its input from four hard-coded sentences and had never read the upstream output at all, from the beginning, while both stages reported success for weeks ^[10]. The reported fix is to record on every artifact which artifact it was made from, so an unread output shows up as a number instead of an absence ^[11].

Language drift is the third-largest cluster at 1048 occurrences ^[12], about 6.9% of type occurrences ^[5]. In a separate model evaluation using the same production contracts, a candidate produced 1,680 characters of confident, well-formatted text for a Japanese review task containing no kana at all; it was entirely Chinese, and it passed both the length and structure checks ^[13]. The operator now gates on kana presence and keeps the writer and reviewer models in different families so a family-wide language failure cannot approve itself ^[14].

The prescription is narrow and cheap: declare the output shape and machine-check it before anything downstream reads it, including minimum length, required and forbidden patterns, expected language, and for numbers whether the figure is labeled measured, target, assumed or estimated ^[15]. The author extracted a zero-dependency, synchronous string checker called honto-contract from that work ^[16], and the second recommendation is to record why a run failed rather than that it failed ^[17]. Note that this is one operator, self-reporting, promoting their own tool.

Two things worth watching. Whether anyone reproduces the 43% split against different contracts, since 15251 occurrences across at most 6768 records means each failure carried at least 2.2 violation types ^[6] and the taxonomy is doing a lot of work. And whether provenance counting, the unread-input check, spreads faster than the eval suites, since it is the one fix here that needs no model at all ^[11].