Скачать или смотреть Refactoring COBOL to Java with Agentic AI with an Iterative Refinement Workflow

Legacy COBOL systems still run critical business infrastructure—but migrating them safely to modern languages is hard, expensive, and risky. In this webinar, OpenHands senior researcher Calvin Smith walks through a practical iterative refinement workflow where one agent refactors COBOL → Java and a second “critic” agent scores the output and generates actionable feedback—looping until it meets a quality threshold.

Links:
Presentation: https://ggl.link/coboltojavawebinar
Blog (COBOL → Java, iterative refinement): https://openhands.dev/blog/20251218-c...
OpenHands Software Agent SDK (GitHub): https://github.com/OpenHands/software...
SDK Docs: https://docs.openhands.dev/sdk
OpenHands Website: https://openhands.dev/
Contact / Get in touch: https://openhands.dev/contact
OpenHands GitHub org: https://github.com/OpenHANDS

Presentation:
00:00 – Welcome & introductions
01:16 – Why COBOL modernization is urgent
02:12 – The real risks of legacy COBOL systems
03:25 – Why COBOL → Java migrations fail
05:02 – Why AI agents struggle with large refactors
06:26 – The core idea: Iterative refinement with AI agents
08:49 – Engineer ↔ Critic agent workflow explained
10:40 – Why this approach works (feedback loops & convergence)
13:16 – Proof of concept results & what was achieved
14:19 – OpenHands Agent SDK overview
17:21 – Tip #1: Traceability from Java back to COBOL
19:17 – Tip #2: Concrete scoring for reliable critics
23:05 – Tip #3: Snapshot tests to validate migrations
25:45 – Scaling to enterprise codebases
27:49 – Large Codebase SDK & dependency-aware parallelism
29:12 – Key takeaways for AI-driven modernization
30:36 – OpenHands programs, resources & next steps

Q&A Session

32:33 – Live Q&A: COBOL migration with AI agents
33:00 – How the critic agent evaluates correctness & quality
35:45 – Reliability with limited COBOL training data
38:00 – Mainframe realities: JCL, DB2, RPG, and legacy systems
40:00 – Long-running agents, context limits & Opus performance
42:00 – Database optimization after modernization
44:30 – Preserving business logic across large codebases
47:00 – Models used, cost considerations & scalability
51:10 – Parallel agents without breaking interfaces
54:55 – Best practices for context engineering at scale
56:25 – Wrap-up & community call invite