AWS's AI Modernization Flow moves 150,000 lines of Java 11 to Java 21 in four days

The AIMF design leans on a simple split: the AI proceeds on implementation inside a unit once verification passes, but stops for human decisions on interface or data-format changes, feature removals and phase boundaries. Whether it stops is judged by pre-set conditions, not by the AI itself. Technical problems get a fixed number of retries before being cut out as a separate task, and unplanned work must be proposed and approved before joining the plan — practical guards against an agent sinking time into a dead end.

The completion report for Roller is the most reusable part. Existing unit tests passed 157 cases, with one legitimately removed along with OAuth 1.0 support; a verification script confirmed zero remaining javax references; browser E2E checks covered user registration, login, blog creation, posting, image upload and comments. Silent failures from security defaults were found and fixed, and a recurrence-detection script was added. How each success criterion was checked is listed item by item, and deferred work is recorded in an ADR with reasons.

Context is kept in an ADR for design decisions, a worklog for tasks and session-context for current state, and a separate agent with isolated context verifies the completion claim once and closes each finding one at a time. That lets work resume after a handover, a day boundary or an abnormal agent exit. The publisher itself notes this is not a magic tool that automatically completes optimal modernization, and that the Roller case is OSS validation — business-system-specific reports and external integrations need separate checks.