Blog of Martin Russmann

The Strangler Strangler Flow (hereafter: the AI Strangler Flow) reimagines legacy system displacement for the agentic AI era. Unlike traditional big-bang rewrites or manual strangler fig patterns, this framework leverages large language models with million-token context windows to comprehend, extract, and modernize legacy systems continuously. Legacy codebases — COBOL, PL/SQL, Java monoliths — can be analyzed far faster than in traditional discovery workflows, often within hours rather than weeks, converted into machine-readable knowledge graphs augmented with structured markdown documents, and strangled incrementally through autonomous AI agents, with human oversight concentrated at critical decision gates. The framework makes two core contributions. The first is a formal governance model — six human intervention gates positioned at irreversible or ambiguous junctures in an otherwise autonomous pipeline — that defines where AI execution must yield to human judgment and why. The second is a self-improving extraction loop, adapted from Karpathy's autoresearch paradigm, in which each completed displacement cycle refines the system's own extraction strategies. The displacement pipeline does not merely execute; it learns, accumulating a meta-knowledge graph of process performance that makes each successive extraction faster, more accurate, and less dependent on human correction. Documentation Innovation: Documentation becomes persistent AI state (knowledge graphs, embeddings, executable specs, and markdown narratives) enabling zero-latency handoffs between AI agents, continuous modernization, and human-readable audit trails. Markdown is not the source of truth — it is a rendered view of the knowledge graph, auto-generated and kept in sync. Humans edit markdown to provide clarifications; AI parses these edits back into the graph.