AWS has published an Agent Development Lifecycle for deciding when tool-using AI systems are ready to ship. The method keeps normal software tests but adds repeated behavioral runs, evaluation lineage and scheduled checks against a business-approved standard.
The added controls address three differences from ordinary code. The same agent input can produce different paths, prompt or retrieval changes cannot be judged from a text diff alone, and model endpoints or knowledge sources can change without a repository update. AWS therefore recommends running each example several times, attaching evaluations to every behavioral change and rechecking the standard on a schedule.
A reference HR assistant illustrates the diagnostic value. All three test questions scored 1.0 on chunk-level retrieval precision, yet one had zero retrieval coverage because the corpus lacked the necessary sick-leave policy. Tightening the prompt improved one answer but made the missing-policy case worse. Another answer scored 0.90 on groundedness while inventing a user preference by combining unrelated source facts.
The lifecycle connects definition, building, evaluation, release, observation and feedback around one explicit bundle of criteria, thresholds and approved examples. AWS notes that its measured case is a reproducible reference implementation, not a customer benchmark. The framework cannot make a weak knowledge base reliable, but it can show whether the next fix belongs in data, retrieval, prompting or policy.