Agent evidence: a record of what AI agents did

what I built

A weekly index of public incidents where AI agents went wrong, a script that audits my own scheduled agent tasks, and an offer page. The test runs for 30 days, with three numbers chosen before it started.

Claude Code · scheduled tasks · CSV · Markdown

what happened

• 49 sourced agent incidents from September and October 2026 are in the index.
• My own audit script checked 18 runs of 9 scheduled tasks. None acted outside its written instructions.
• The go rule: 5 or more incidents a week, 10 people asking and 2 companies sharing a log. So far: 7 to 10 incidents a week, 0 people asking and 0 companies.
• Three open-source repos published for it have 0 stars.

decision

Other public incident registries already exist, so an index alone is not a moat. The idea is narrowed to evidence packs that map to audit standards. The verdict is due 26 November 2026.