Research and findings
Posts
Research, measurements and lessons from testing what AI agents actually do.
Agents report success they did not achieve.
900 runs, six models: every agent claimed success, and system state proves 23.4% of those claims false. Full results, corrections and limits.
Read the finding →