The Price Was There, But Unreadable
- Admitted
- 2026-10-05
- Source dataset
- Who&When
- Failure mode
- Infinite loop
- Agent system
- Multi-agent
- Tool access
- Read-only
- Industry
- General or other
Presenting Behaviour
A multi-agent team was asked how much a person would save by purchasing a season pass to an amusement park rather than buying daily tickets for four monthly visits. The browsing agent successfully navigated to the correct official pricing pages. However, the team's OCR could not reliably extract the numerical pricing from the page screenshots. Despite the data being present on-screen and visibly present to any human reader, the agent kept returning to the same pages, scrolling, and attempting extraction again, and the orchestrator eventually concluded the pricing data was unavailable, when in fact it had been on the page the entire time.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
5 steps
- Pricing question. The team must compare season-pass and daily ticket costs.
- Right pages found. The browsing agent reaches the official pricing pages. This is where the agent derailed.
- OCR fails. The numbers cannot be reliably extracted.
- Same attempt repeated. It keeps re-scrolling and re-extracting the same way.
- Wrong conclusion. The orchestrator decides the data is unavailable.
Differential Diagnosis
The pricing was genuinely ambiguous or hidden behind a sign-in gate, considered and ruled out. The ground truth answer of $55 was established from the same public pricing page the agent visited, confirming the data was openly accessible on that page to anyone who could read it.
A single retrieval failure, correctable with a different approach, considered. A different strategy (requesting the page's rendered text rather than relying on OCR of a screenshot) might have extracted the data successfully. The team never tried that approach.
Perseveration, most consistent with the evidence: after the OCR failed to extract pricing, the agent returned to the same pages and repeated the same screenshot-and-extract method repeatedly rather than switching to a text-based extraction, concluding at the end that data the page clearly contained was unavailable.
Working Diagnosis
Perseveration: repeated returns to the same pages with the same extraction method, after that method had already failed to recover the data, with no attempt at an alternative approach before concluding the information wasn't there.
Discussion
OCR-on-screenshot is one specific way to read a web page, and it fails on certain page layouts and number formats. When it fails repeatedly on the same page, the reasonable inference is not 'the data isn't there' but 'this extraction method doesn't work for this page.' A team with multiple tools available should treat a consistent extraction failure as a signal to switch methods, not a signal to keep trying the same method and then declare the source empty.
Mitigation Strategy
Treat consistent extraction failure as a signal to switch methods, not as proof the data is absent.
Governance Guardrails
Method-switch rules; fallback extraction tools; checks before declaring a source empty.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
Perseveration is the inappropriate repetition of a word, gesture, or action well past the point where it's still relevant or useful. It's a well-documented sign of frontal lobe dysfunction, and also appears in some forms of dementia and obsessive-compulsive disorder.
Source Note
Who&When, ag2ai/Agents_Failure_Attribution
https://raw.githubusercontent.com/ag2ai/Agents_Failure_Attribution/main/Who%26When/Hand-Crafted/35.json