RPG Code Generation And The Agentic Future Of IBM i

June 2, 2025 Alex Woodie

IBM is making progress with Watsonx Code Assistant for i, which is on track to become available as a public preview next month. And while IBM perhaps is running into some challenges with generating RPG code, the long-term potential of agentic AI on IBM i is brighter than ever.

As we previously reported, IBM is now taking names for the folks who will be given access to the public preview of WCA for i, which you can sign up for here. According to IBM i CTO Steve Will, the preview will deliver RPG code understanding capabilities, which is the first batch of AI co-pilot features targeted for this product.

Will, who is also an IBM Distinguished Engineer and chief architect for IBM i, elaborated on the state of the WCA for i project and the challenges that his development team has faced during an interview with IT Jungle at COMMON’s POWERUp 2025 conference in Anaheim last month.

“I have grown to understand better all the different things people wanted from it, so in that sense, it’s expanded,” Will said. “But I knew ahead of time I didn’t know enough. We have never built anything like this before. And we knew we needed to learn how to train a model to do these things.”

IBM largely is sticking to the timeline of delivery of features that Will unveiled when he first announced the coding co-pilot project one year ago at the POWERUp 2024 conference in Fort Worth, Texas.

The first goal was delivery of code explanation capabilities in WCA for i. Code explanation consists of four subcomponents, including detailed explanation for a programmer, a business-level summary, an explanation of usage, and then documentation that can be inserted into source code. The private preview release of WCA for i that will become available in July will deliver these code explanation features for VS Code; support for Rational Developer for i (RDi) is not currently in the plans.

Will wants WCA for i to be fairly good at code explanation before moving on to the other two components, including test case generation and RPG code generation.

“That’s one of the reasons we’re focusing on explain right now, and it’s actually going better than we thought,” Will said. “But it’s going to take a little while before we get to some of that generate stuff.”

Will is hopeful that his IBM team will deliver some RPG code generation capabilities in WCA for i by the end of year, when the product is currently slated to become generally available. But that is still up in the air, he said, indicating that that component may be added to the product at a later time.

“I hope that we’ll have some generate thing by the end of the year, but whether it’s by the GA of the product, I’m not exactly sure,” Will said. “It might take a couple of months more. But we’re not going to wait for another year and then do another big GA. We’re just going to roll out these things as we train the model and create a user interface to use that.”

WCA for i has been modeled off other WCA products, including Watsonx Code Assistant for Z, which the mainframe group developed to understand and generate COBOL, the primary language used on the System Z. While RPG and COBOL are both very old procedural languages, it turns out that RPG and COBOL have significant differences in structure. That has contributed to the learning curve for Will’s group.

Will’s group of IBM developers aren’t the only RPG experts who have taken a gander at building a coding co-pilot for RPG. Fresche Solutions also experimented with training a LLM to understand RPG in laboratory research. The experiments did not go well, according to Fresche chief technology officer John Clark, and the company opted against training an LLM directly on RPG with the recent update to X-Analysis, which utilizes the native code-understanding of third-party models like Anthropic’s Claude, widely considered one of the best LLMs for coding tasks in the industry.

POWERUp presentation by IBM i CTO Steve Will.

“We really struggled with RPG,” Clark told IT Jungle in a recent interview. “The challenge is always in the aspects of RPG which may be implicit. Think of the RPG cycle. That’s not easy for these LLMs to understand. Think about some of the key characteristics people use all the time, like data areas. LLMs struggle with those kinds of concepts and what’s going on with that.”

Relayed Clark’s words, Will nodded in recognition of these challenges during the briefing at Disneyland Hotel. “It doesn’t surprise me that somebody else who’s working with RPG would say that,” Will said. “We use components from the Watson Code Assist core, and the Watson Code Assist core knows a lot of procedural languages pretty well. RPG is different enough that there were things we had to do differently and train differently.”

IBM has fed more than 10 million lines of RPG code into the Granite models, Edmund Reinhardt, the IBM i application development tools lead for IBM, said during a POWERUp 2025 session. The company needs a lot more RPG code – especially “old and ugly” RPG II and RPG III code, the kind of stuff that powers a good chunk of the legacy IBM i applications that IBM so desperately wants customers to modernize.

According to Will, the RPG understanding aspect of WCA for i is going better than expected. While the code generation may not be ready by the end of the year, Will insists that it’s not a hurdle.

“We didn’t know what it was going to take to do this, so I’m glad we have so much internal knowledge of how RPG is built and used inside, and that we’re able to lean on the private preview participants who know so much about how it’s been used so that we can take advantage of it,” he said.

“I really want to get this message out: We need more people to donate, contribute older RPG. Because not only is it fixed-format, which we have to train the thing to do, but there were things in previous versions of RPG – F specs and other specs that I don’t even know what they were – that don’t exist in other languages. And we’re going to have to train the model to understand what that means so that it can explain it, at the very least.”

RPG is the current focus of WCA for i, but IBM will soon be looking at supporting other IBM i languages with it. That includes SQL, CL, and COBOL, which is similar but different than COBOL on the mainframe.

“We have not yet validated how good the model is at explaining SQL. We expect it will be pretty good because the Granite model in general is, but we haven’t tested it and we think we’re going to need help,” Will said. “We know we’re going to have to train a model eventually on CL if we want to do CL. It’s not just going to know it on its own. And COBOL on IBM i is different than COBOL on Z, so we’re going to have to have some COBOL on IBM i examples.”

WCA for i looks to be a formidable tool for application modernization tasks. The capability to understand ancient RPG code that hasn’t been documented and re-generate it as modern free-format ILE RPG code will be nothing short of a godsend for the thousands of IBM i applications who don’t see a path forward with their legacy IBM i applications. Similarly, AI can help customers modernize older DDS data definitions to use newer SQL-based DDLs.

Will also sees a use for WCA for i by modernization tool vendors.

“Between now and GA, we have to talk to the modernization vendors and give them a chance to play with this,” Will said. “Do they think they could use it right out of the gate? Do they need some API level support? I don’t know. I agree that there is room for people who are not us – that is, not IBMers – to do specific things that we’re not doing, and I’m hoping that we can make it so that they can take advantage of this stuff. It’ll still have to be built on top of our stuff.”

Once the RPG, SQL, CL, and COBOL coding copilot is built, IBM seems destined to set its sights on the Next Big Thing: agentic AI. Will and his business architects are already looking into whether and how IBM i would support Model Context Protocol (MCP), an emerging standard that Anthropic released barely six months ago to provide a repeatable way for AI applications and AI tools to access data.

The potential to automate a range of administrative tasks on IBM i using AI agents is a tantalizing one, indeed–and thanks to the rapid pace of AI technological development, it may come to fruition much sooner than one may expect.

“My vision for this is that we know already there are MCP clients out there that can consume the tools built on MCP, and we know that it is very trivial to turn all of our SQL services, our IBM i services, into agents,” Will said. “What we have to do is figure out what does it take for us to build the server, to expose all of that? And then again, where can a partner put their stuff on top of that server? Where will it make sense?

“It’s sort of like what we did when we built the SQL services, the IBM i services,” Will continued. “We did it so that Navigator and ACS could have simple ways to do all of these things through a natural interface, and then we made them all available to everybody, not just for us. I want to do the same sort of thing with the agentic stuff, make the information that everybody should have access to available in a common way.”

As the old saying goes, no good deed goes unpunished. Will may have had relatively narrow goals when he started the WCA for i project a year ago. As the world of AI agents blossoms in a technological spring, it appears we may be on the verge of a full-blown AI-ification of the entire IBM i platform. Will said IBM will likely start with one or two use cases for agentic AI, but obviously it won’t stop there.

The potential to have a fully automated business system has never been so close. And to think it could come out of the IBM Rochester lab first. Makes you wonder: What would Thomas Watson Jr. think?

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search