import random
from agent_reliability_toolkit.live_eval import bisect_failures

def live_bisect(agent, eval_cases, max_failures=5):
    actual_failures = []
    for case in random.sample(eval_cases, len(eval_cases)):
        result = agent(case['input'])
        if not result['success']:
            actual_failures.append((case, result))
            if len(actual_failures) >= max_failures:
                break
    return actual_failures

# Archive these for manual inspection and regression