Japanese bank: 4M lines of COBOL, 3 AI agents, 60% migrated in 14 months.

4 million lines of COBOL, 12 people to maintain it, 9 retiring in 3 years. That's not technical debt, it's a cliff.

The context

Japanese regional bank, 1,800 employees, Tokyo hub. Core banking system written in COBOL since 1987, two generations of mainframes behind. Patchy documentation, opaque dependencies, business rules scattered across 40-year-old copybooks.

The problem

• 4M lines of COBOL, 9 of 12 maintainers retiring in 3 years
• Documentation 40% obsolete, business rules untracked
• Release cycle: 14 weeks for a minor change
• No automated tests on the core, 3 weeks of manual validation
• Annual maintenance cost: 4.2M EUR (mainframe licenses + outsourcing)

The approach

Abbeal multi-agent method: three specialized AI agent roles, supervised by 8 human engineers. The Archaeologist documents existing code, the Architect recomposes into Java/Kotlin services by bounded contexts, the Cleaner removes dead code after equivalence validation.

The principles

• Migration by bounded contexts (DDD), not by technical modules
• Equivalence tests: each migrated service runs in parallel with COBOL for 60 days
• Strangler pattern: progressive routing via API gateway
• Formalized knowledge transfer: each migration produces a runbook readable by juniors
• Mandatory human veto before any agent-produced merge

The stack

• COBOL source (IBM z/OS mainframe), legacy copybooks
• Java 21, Kotlin 2.0, Spring Boot 3
• AWS Bedrock (Claude Sonnet) for Archaeologist/Architect/Cleaner agents
• OpenSearch for COBOL corpus indexing and knowledge graph
• Camunda for business workflow orchestration

The results

1. 60% of COBOL estate migrated in 14 months (2.4M lines)
2. Release cycle: 14 weeks to 5 days on migrated services
3. Maintenance team reduced: 12 to 4 people (successful knowledge transfer)
4. Annual maintenance cost: -38% starting month 14
5. Zero production incident attributable to the migration

What we learned

AI agents are excellent at documenting and proposing, bad at deciding alone. The 1 human / 3 agents ratio holds over time, beyond that quality drops. Parallel equivalence tests are expensive on infrastructure but non-negotiable: three subtle bugs detected only there. To redo: start with the least critical bounded context, not the simplest. Complexity comes from dependencies, not from the code itself.