A two-night learning trace that is explicitly illustrative

Source rechecked October 24, 2026: Nameet Dutia and Imen Grida Ben Yahya, “Building self-learning agents for intelligent Network Operations with Kiro Crew”, AWS Industries, September 30, 2026. The authors explicitly call the FN-114→FN-208 overnight sequence a sample illustrative scenario for a cable network that an operator might implement; its dates, fiber nodes, decisions and later success are not observed network incidents.

The proposed loop stores an engineer's correction as a conditional diagnostic test rather than yesterday's answer: compare upstream with downstream signal degradation before recommending amplifier replacement, after checking plant design and amplifier family. A later fictive session checks applicability, uses the same test on different signals, and stages a repeatable skill for human approval. Lessons are persistent but have no native expiration/review date; operator curation must prevent stale rules. Dynamic skill creation is off by default and generated skills await approval. This is an explanation of a possible handoff and its trust boundary, not proof it improved field reliability.

Inspectable engineering, different endpoint. The Kiro Crew maintainer changelog, checked October 24, describes a short skill index and on-demand search, always/on-topic and repository lesson scopes, a malformed lesson blocking chats and scheduled runs, and preserving an existing skill if its replacement fails. These are concrete maintenance changes; the changelog does not show a corrected lesson retrieved, acted on, and independently verified at a later incident. See Kiro's separate August model-serving operator case for an actual per-cell alarm correction without later reuse trace and the sandbox old/new comparison for measured discovery loss.