The periodic drop
The State of Agent Risk.
Everything the registry knows, published in full with the methodology beside it so it can be quoted without being misquoted.
The corpus, by quarter.
The series rather than a growth multiple. A multiple over two thin quarters is a decimal wearing the costume of a trend — the counts are here and you can decide.
What goes wrong, across everything
And the thing it cannot say: how often this happens.
There is no denominator. Nobody publishes how many agent deployments exist, so no rate can be computed from this corpus. Any incident rate you see quoted for AI agents — including one derived from this dataset — was produced by inventing a denominator.
Methodology, in full
These figures count incident REPORTS filed to agitome's public registry, plus the small subset that has completed adjudication against signed loss evidence. They are not a survey and not a sample: they describe what has been reported, by whoever chose to report it. Three limits follow, and none of them is fixable by counting harder. (1) There is no denominator — nobody reports how many agents are deployed, so no rate per deployment can be computed and none is published here. (2) Selection bias is severe: a framework with more users, more visibility, or a more engaged community will accumulate more reports regardless of how it actually behaves. (3) Withdrawn and disputed reports are excluded from every statistic but remain publicly visible, so the counts here are smaller than the catalogue. Any claim of the form 'X% of agents fail' cannot be supported by this dataset.
- Verified claims
- 0 subjects have a party present to answer
- Open disputes
- 0
- Excluded from statistics
- 0 withdrawn or contested
- Generated
- 2026-08-13T13:51:59.002348Z