The vibe is the last 10%. The expertise is the other 90.
Vibe coding is real. It works.
You describe what you want. The AI writes the code. You review, adjust, iterate. For web applications, Python scripts, React components – the productivity gains are genuine and significant.
And the results are interesting.
Vibe coding is the practice of directing an AI to write code through natural language – describing intent rather than implementation. You provide the what. The AI provides the how.
The concept works because modern AI models have been trained on enormous volumes of code. They have seen enough Python, JavaScript, Java, and Go to produce syntactically correct, often functionally correct output for well-understood problems.
The key phrase is "well-understood problems." Vibe coding works when the AI understands the domain it is coding in.
This is where mainframe gets interesting.
Ask an AI to write a Python function that reads a CSV file and calculates totals. It will produce working code on the first attempt. The domain is saturated – billions of examples exist in training data.
Ask an AI to write a COBOL program that reads a VSAM KSDS file, processes records with packed decimal arithmetic, handles SQLCODE -811 correctly, and writes output with the correct RECFM=FB format for the downstream job.
Something different happens.
The AI will produce code that looks correct. It will compile. It may even pass basic tests. But subtle errors will hide in the places that only experienced z/OS developers know to look:
These are not random errors. They are systematic errors that come from training on limited mainframe examples, from the complexity of the z/OS execution environment, and from the gap between what the code says and what the system does.
Here is the critical insight: vibe coding does not replace technical expertise. It amplifies it.
A developer who deeply understands COBOL, JCL, VSAM, CICS, and DB2 uses vibe coding as a productivity accelerator. They describe the intent, the AI produces a first draft, and the developer applies their expertise to review, identify the subtle errors, and correct them before they reach production.
The AI handles the mechanical work. The developer handles the judgment.
The result is genuine productivity gain – the same developer produces more, faster, without sacrificing quality.
They cannot identify the packed decimal overflow risk. They cannot spot the CICS COMMAREA assumption. They cannot read the JCL and know that DISP=NEW will fail on a re-run.
They will see code that looks right, passes initial tests, and reaches production carrying problems that only surface under specific conditions – the exact conditions that do not appear in a test environment.
This is not a criticism of AI tools. It is a statement about the nature of expertise and where it is irreplaceable.
Every development platform has a knowledge floor – a minimum level of expertise below which vibe coding produces unreliable results. On web platforms, that floor is relatively low. The feedback loop is fast, errors are visible, and the cost of failure is bounded.
On mainframe, the floor is higher. The reasons are structural.
The execution environment is invisible to the AI. JCL, VTOC, catalog entries, JES2 routing, RACF profiles – these are not in the source code. The AI writes the program. It cannot see the environment the program runs in.
The error surface is large. A COBOL program interacts with datasets, databases, transactions, and other programs. An error in one interaction may not surface until a specific combination of conditions occurs in production.
The cost of failure is high. These systems process payroll, settlements, insurance claims, and financial transactions. A subtle error that reaches production does not just cause a bug report. It causes a production incident.
The knowledge required to work safely in this environment is not acquired in weeks or months. It is acquired over years of working with real systems, real data, and real production incidents.
Vibe coding on mainframe works. It works well when the developer using it understands what they are asking for and can verify what they received.
It fails – quietly and dangerously – when that understanding is absent.
The organizations deploying AI coding tools on mainframe without investing in the technical foundation are not accelerating development. They are accelerating the production of code that looks correct, tests adequately, and behaves unexpectedly.
The answer is not to avoid AI tools on mainframe. The answer is to treat technical expertise as the prerequisite, not the optional extra. Train developers on the platform first. Then give them AI tools to accelerate the work they now understand.
The AI amplifies what the developer knows.
It cannot substitute for what the developer does not.
The developers who get the most from AI coding tools on z/OS are the ones who have spent years learning why things work – not just that they do.
The vibe is the last 10%. The expertise is the other 90.
Related reading: What AI Can and Cannot Do with COBOL Source Code · Why AI Pilots Succeed and Productions Fail · We Are Building for the Enterprises That Stay