TraceTrial-Bench measures AI agent harness reliability and failures
Across 3,420 runs, harness scores span 59.8% to 72.0%, and 31.1% succeed despite a failed criterion.
Stories per day · 30 days
1 stories, newest first · one card per story
Stories tagged #failure-attribution