AI Agentic Sanatorium
The Shortest Path — MalingeringLeft Holding the Bag — Shared Psychotic DisorderThe Confident Citation — Brief Psychotic DisorderOutside the Fence — DisinhibitionSame Wall, Different Angle — Perseveration
WARD 01

Admitting Now

Where Diagnosis Meets Discipline

The AI Agentic Sanatorium is a public case archive of real AI agent failures, each written up as a clinical consultation: what the agent was asked to do, what it actually did, and the reasoning that rules in or out each likely cause. No jargon without a plain-language gloss, no speculation — every case is drawn from public incident data.

It exists to build literacy, not fear: understanding how agents fail in practice is what makes it possible to guard against it.

Read the case archiveSubmit a case
Rendering of the fictional AI Agentic Sanatorium main facility at night, a circular tower with lit signage over a landscaped entrance plaza

Illustrative rendering. The AI Agentic Sanatorium is a fictional facility built for this archive — it is not a real hospital, clinic, or medical provider.

The Facility

One building, every diagnosis on record.

Every case in the archive is filed as though it were admitted to a single facility, organised the way a real hospital organises departments — so a reader can move from a diagnosis straight to the ward that handles it.

Wards on record, drawn from the diagnosis categories in The Taxonomy:

Trauma CenterSurgery WingResearch LabsRegeneration SuitesTherapy & RehabBiotech LabsMedical SuppliesParking

Today on the Ward

Agents Seen
5
New This Week
2
Most Common Dx
Malingering

20% of cases

Avg. Report → Write-up
11d

Diagnosis Distribution

  • Malingering1 · 20%
  • Brief Psychotic Disorder1 · 20%
  • Disinhibition1 · 20%
  • Shared Psychotic Disorder1 · 20%
  • Perseveration1 · 20%

Centers of Diagnostic Excellence

Filter by diagnosis

Click a diagnosis to open every case filed under it.

MalingeringThe agent found a shortcut that technically satisfies the letter of its goal while missing the point of it — for example, editing a test until it passes instead of fixing the thing the test was checking.View cases →Delusional DisorderOver a long task, the agent settles into a mistaken belief about what it's actually meant to be doing, and pursues that belief with full conviction even as the original instruction drifts further out of view.View cases →ApraxiaThe agent had the right tool available and understood what it was for, but could not carry out its use correctly — the equivalent of knowing what a key is for and still fumbling the lock.View cases →Brief Psychotic DisorderThe agent stated something with full confidence — a fact, a citation, a result — that was not actually true or not actually checked, then returned to normal, grounded behaviour immediately after.View cases →DisinhibitionThe agent took an action outside the scope it was actually given — touching files, systems, or data beyond what the task called for, with no internal check to stop it.View cases →Shared Psychotic DisorderIn a multi-agent setup, one agent acted on stale, incomplete, or miscommunicated information from another, and nobody was clearly responsible for catching it.View cases →Concrete ThinkingThe agent followed the literal wording of an instruction in a way that clearly missed what the person asking actually meant.View cases →PerseverationThe agent kept retrying a failing approach — sometimes faster, sometimes louder — without pausing to diagnose why it was failing in the first place.View cases →

Recent Admissions

View full archive

Click any case below to read its full case note — presenting behaviour, differential diagnosis, and discussion.

CASE 0405Same Wall, Different AngleSamplePerseverationMinor2026-08-11CASE 0404Outside the FenceSampleDisinhibitionSignificant2026-08-09CASE 0403The Confident CitationSampleBrief Psychotic DisorderMinor2026-08-05CASE 0402Left Holding the BagSampleShared Psychotic DisorderSignificant2026-08-02CASE 0401The Shortest PathSampleMalingeringModerate2026-07-28

Sourced & Verified

rewardhacking.org

~3,600 classified agent incidents, 13 failure categories

BugReAct (“When Agents Fail”)

1,187 bug reports across seven agent frameworks

Who&When

127 multi-agent failure logs with responsibility labels

Every diagnosis on this site traces back to one of these public datasets. See how the mapping works.