As I discussed in my previous deep-dive (see here), IBM’s watsonx Code Assistant for Z (WCA4Z) has been on quite a journey. Now, navigating through the rolling updates of version 2.8 (with the latest 2.8.20 patch dropping just days ago in late February 2026), we are looking at the culmination of this release cycle.
This is exactly what we expected: the final major iteration for this product before it is succeeded by the highly anticipated Project Bob. And while the release announcements often read a bit like a game of buzzword bingo, digging beneath the marketing veneer reveals some genuinely useful, enterprise grade features that have been heavily fleshed out over the last two months.
Here is my take on what the WCA4Z 2.8.x series brings to the table, and where the reality check is still needed.
True Architectural Reasoning
It’s easy to look at the current AI landscape and dismiss this update as just “another agentic wrapper”, like throwing RooCode and GPT at a mainframe. But that comparison falls short. The real magic in v2.8 lies in Z Understand. This isn’t just an LLM wrapper with a few MCP (Model Context Protocol) tools attached; it acts as a deep metadata analyzer for your program sources. It allows the system to capture cross-application dependencies for programs that don’t even reside in the same workspace.
Recent patches have supercharged this. The WCA4Z chat now actively uses MCP to pull this deep understand data, allowing architets to initiate Business Rule Discovery directly from an agentic chat. The system can query structural metadata, perform dependency analysis, and understand the impact across the entire application architecture, all without overloading the LLM context window with thousands of tokens.
To orchestrate this, IBM introduced a new multi-agent setup:
- Z Orchestrate Agent: The manager that coordinates the workflow and queries the RAG (Retrieval-Augmented Generation) system.
- Z Architect Agent: The researcher that retrieves the Z Understand metadata, performs impact analysis, and feeds this structured context forward.
- Z Code Agent: The executor responsible for actual code generation and acting as your refactoring assistant.
In my initial look at the December 2.8 release, the Z Transform and Validate Agent for COBOL-to-Java was noticeably absent, leaving me highly skeptical about the state of automated Java transformation.
What should I say: I spoke too soon.
With the 2.8.20 release just days ago, the chat now officially supports the transformation of COBOL and PL/I programs to Java. They also rolled out support in 2.8.10 for IMS programs using CBLTDLI and EXEC DLI interfaces (though it explicitly lacks support for multiple segments or IMS service calls).
Given the historical difficulty of automated mainframe-to-Java conversions, I’m keeping my skeptical hat on, but the tools are officially in our hands to test.
Let’s see how it handles messy, decades-old spaghetti code!
Explain, Document, and Optimize: The Quality-of-Life Upgrades
Beyond the architectural agents, the continuous updates through January and February brought also some quality-of-life improvements that we developers will use daily: The Code Assistant now automatically expands copybooks and INCLUDES when generating documentation or explaining code in the chat (for small programs). So, we now get even more context for these tasks.
Z Code Scan
While it wasn’t the star of the executive blog, Z Code Scan is a feature developers will love. It provides automatic coding standards compliance based on your enterprise-specific rules. You feed the tool your internal coding guidelines, and it provides both findings and resolution guidance.
Even better, the generative agents actively use these rules when writing new code. Sure, other linting tools exist, but having a seamless way to tell your AI agent exactly how your company writes code is incredibly valuable.
Reliability and the Multi-Model Shift
Now for the skepticism. Reliability remains the ultimate killer feature. Until we validate these new capabilities with 100% reliability in messy, real-world scenarios, they remain a highly promising vision rather than a daily guarantee.
There is also a fascinating pivot in the engine room. Every new agentic feature in v2.8 runs on Mistral models, while IBM’s Granite models remain available for non-agentic tasks. Rather than a shortcoming, this suggests IBM is adopting a highly pragmatic, “best-fit” multi-model approach, leveraging Mistral’s specific strengths for complex orchestration and heavy-duty agentic function calling, while keeping Granite for tasks where it already excels. It’s a smart move that prioritizes the best tool for the job.
Final Thoughts
WCA4Z 2.8.x is more than just another AI agent; the integration with Z Understand and its enterprise-scale architectural reasoning is a major step forward. It effectively sets the stage for the highly anticipated transition to Project Bob, giving us a strong preview of the deeper, multi-agent workflows that will soon be natively baked into the IDE.
It is absolutely worth a serious technical evaluation (a Proof of Value, not just a Proof of Concept) today. But as always, you have to actively look for the gaps between the marketing promises and your operational reality as you prepare your modernization strategy for the next generation of mainframe AI.
