In previous posts titled Anthropic code or Hypethropic code? and Anthropic code: No Revenues Plan as Immaturity indicator I was very skeptical about current Anthropic code capabilities.
This post is about the unrealistic dream of transforming Legacy Mainframe COBOL Systems to modern systems by using Claude code COBOL tool.
More than 50 years ago I was a COBOL Programmer.
Afterwards I was a Systems Programmer and a CTO.
I participated in Mainframe systems architecture changes, replacement and maintenance, as well as Mainframes migration projects and Modernization projects.
The main issues are not pure coding issues. Some of the major obstacles will be dwelled upon in the following paragraphs.
Functionality and flow of the System
Usually nobody knows and understands it completely.
Claude code COBOL tool will create System which is not downward compatible.
The systems were written decades ago and if there is documentation, probably it is not updated.
The original programmers not necessarily available.
Those who maintain the Systems do not understand the System and the code as well as the original programmers understood it.
Large Unstructured Silo Systems
Many Mainframe systems are Large Silo systems.
The systems are not Component Based or Service Oriented. It is not possible to transparently replace a component by modern code such as code created by Claude Code.
Data and Systems Integration
The APIs between systems are not standard modern API's
Could Claud Code build them correctly?
Other less known Programming Languages
The Mainframe applications include code written in other programming languages.
Among them are Assembler, Obsolete programming languages such as Mark4, UFO and other languages such as PL/I, Software AG's Natural.
Bugs and Debugging
Even in cases of AI tools writing simple applications there are many bugs.
Expect for more bugs in large Mission Critical Mainframes systems.
Performance Issues
Even if the code will be fully compatible to functionality requirements, bad Performance could limit its usage.
Is the performance of code created by Claude Code good enough?
Large Systems performance requirements are different from small simple single user systems requirements.
Maintenance
Most of the resources spend during Application Life Cycle are for Maintenance and Application Changes (Estimated as 70% or 80% in Large old Silo Systems).
How difficult is Claude Code COBOL Maintenance?
Reducing Migration Risks
In many cases Migration Projects are performed in steps in order to reduce Risks.
Each step includes Migration of part of the System.
However, the mixed new and old code System should be integrated in order to perform its task as it did before the partial migration.
The integration is maintained by writing and executing temporary code, such as Bridges and APIs.
After verification correct operation of the the part of the system which was migrated together with other parts the next step should be executed.
The temporary APis and Bridges should be adapted for the next step.
Is Claude Code Capable of creating those Bridges and API's?
Conclusion
Currently, Claude Code is far from being able to create COBOL Mainframe Systems or other COBOL Systems such as UNIX Microfocus Cobol Systems.
My Take on Migration from Mainframe
Migration Processes from old Mainframe Systems are difficult and expensive.
The Justification of such Migration projects is not easy.
The top Management could identify many Risks and no direct Business benefits.
The best scenario is new system which is performing the Business Functionality and Processes as good as the previous system performed without any addition of Functionality and Processes e.g. no Business Value.
In many cases the best way is System Modernization i.e. gradually transforming the Silo System to Services.
Services could be Reused and moved to other platforms or recoded in modern programming languages.
Previous promises of migrating Core Banking Systems of Large Banks to other platforms on the Public Cloud failed.
These promises were not realized due to some reasons similar to the reasons that Claude Code is not capable to accomplish its promises.
For example, read Public Cloud Core Banking: Hype or Reality?
