Agent redundancy fails a 90% co-failure test

Two instances of the same model co-failed on 90.0% of the missions on which either failed in a preregistered study of 18,000 two-agent handoffs. The Agent Behavioral Contracts II paper says swapping in a different model reduced the association in all six tested contrasts, while changing vendors without changing the model produced no improvement. Teams adding a second agent for resilience should diversify the model failure modes, not just double the same dependency; the result sharpens the archive’s harness-as-enterprise-moat argument.