Anthropic said in February that Claude can now explore and analyse legacy COBOL on its own. Microsoft, GitHub, Sourcegraph, and IBM all shipped agent-based tooling for the same job this year. The framing in most of the coverage is migration: point an agent at your mainframe, get modern Java out the other side.
That framing is wrong, and it will burn teams that believe it.
The problem was never the syntax
There are roughly 220 billion lines of COBOL still running in banks, insurers, government, and healthcare. The scary number is not the volume. It is the age of the people who understand it. The average COBOL programmer is around 55, and about 10% of that workforce retires every year. The code still runs. The context that explains why it runs the way it does is walking out the door.
An agent that can read COBOL does not fix that on its own. It does something more useful first. It can read a 40-year-old program nobody on your team wants to open, trace the data flow, and tell you in plain English what it actually does. That is reverse engineering, not translation. And it is the part of modernisation that has always been the real bottleneck.
Documentation is the deliverable
The teams getting value out of these tools right now are not asking the agent to rewrite anything. They are asking it to explain. Feed it a batch of programs and copybooks and have it produce: what each module does, which business rules are buried in it, what talks to what, and where the surprises are. The dead code. The undocumented edge case that only fires on the last business day of the quarter.
That output is worth more than a speculative Java port, because it is verifiable. A domain expert can read a business-rule summary and say "yes" or "that is not right." Nobody can eyeball 200,000 lines of machine-translated Java and vouch for it.
Anthropic's own guidance and every serious writeup this year lands on the same caveat: keep humans in the loop, validate everything, and treat full automation as years away, not months. Each COBOL estate is its own strange animal.
How to actually use this
If you are sitting on legacy systems, the practical move is a sequence, not a leap.
- Start by having agents generate documentation and dependency maps for the code you understand least. That is pure upside with no production risk.
- Use those maps to find the seams. Modernise the pieces you can isolate and test. Leave the load-bearing core alone until you can prove behaviour.
- Build a regression suite from real historical inputs before you change a line. If you cannot replay yesterday's transactions against the new code, you are not migrating. You are gambling.
- Treat the agent's translation as a first draft that a person owns, never as the finished system.
The story the vendors want to tell is "AI rewrites your mainframe." The story that holds up is quieter. AI finally gives you a way to read what you inherited, before you decide what to do with it. Recover the knowledge first. The rewrite, if it happens at all, comes later and safer.
We''re here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you''re looking to build something, get in contact with us today!