When Agents Meet Legacy: Never Change a Running System

I've spent a decade debugging undocumented C code, hand-written C++ Makefiles, and grown Java monoliths, knowing exactly which files will break production. Left "allowed to fail" CI/CD pipelines because fixing root causes is harder than ignoring them. I was curious: Can AI Agents help with modernization, and where would they still fail? We'll live-code: untangling complex architectures in C and C++, and if Rust is the answer to everything, breaking Java monoliths to Kubernetes, where agents propose but humans validate, and fixing CI/CD burning budgets when culture resists tooling. The Tanuki IoT Platform environment presents more challenges when agents meet COBOL. What I've learned: agents can't replace judgment on "should we change this?" You'll leave knowing where they help triage, where they fail on business context and experience, and which human strategies, like debugging, you still need. Agents are powerful guides, but the messy decisions are still yours.