Missing context kills mainframe rewrites, not bad code generation. In an AI Engineer session, Neo4j's Stephen Chin and Blitzy's Neeraj Deshmukh mapped a 35-million-line mainframe in four days. I buy the diagnosis.
Models advertise million-token windows, Chin said, but usable context is 200,000 tokens or less. Deshmukh put effective context at 20 to 30 percent, or about 200,000 to 300,000 lines of code. Vector search gives similarity and relational search through MCP gives single-hop facts, Chin argued, while a graph preserves multi-hop relationships.
The interest demo makes it concrete. Deshmukh walked through an AWS card demo where a batch job calls an interest program, pulls copybooks and layouts, and writes VSAM data. Money output depends on inputs, jobs and account updates, not just the formula. Telling AI to convert the COBOL formula to Java finds the formula, Deshmukh said, but misses record precision, account updates, transaction output and batch contracts. COBOL rounding differs from Java rounding, he said, so clean Java code drifted by one cent until parity tests caught it.
That parity claim carries the customer proof. A transpiled job produced output byte for byte identical to the golden mainframe baseline at a Fortune 100 insurer, Deshmukh said. A tier one bank monolith with stored procedures and cross database writes became six capability modules in PostgreSQL, Java, TypeScript and Python in 18 days, he said, versus an 18 engineer week plan, with more than 4,000 tests and modern code less than half the old size. A 25-million-line policy admin platform finished nine times faster than IDE assisted work, he said, with 85 percent autonomously completed and almost 90 percent reaching production.
The enterprise framing matters here. Blitzy announced a collaboration with Microsoft to accelerate legacy modernization at scale, according to the announcement. That pact and TFIR reporting that Blitzy positions as an enterprise platform with forward-deployed engineers, not an IDE assistant, per TFIR, point to the same bet: shared plans and full estate context beat solo agents. The concession is fair. All parity and speed results were vendor stated on stage, Deshmukh said human approval gates every key step, and the team spends more compute on testing than on code generation.
That testing bill is why the claim survives. If you cannot trace every edge, revision and symbol, you cannot promise the ledger still balances. If you cannot prove byte for byte parity, do not let AI touch the ledger.
