The Mainframe Modernization Crossroads: Evaluating Proprietary GenAI Abstraction vs. Open DevOps Architecture

Author: Nivedita Mukherjee Category: Enterprise Architecture / Mainframe DevOps Abstract Modern enterprise IT stands at a critical infrastructure crossroads. Decades of core banking, insurance, and governmental systems rely on z/OS environments that face an escalating crisis: a rapidly retiring talent pool and rigid release cycles. To bridge this gap, organizations are presented with two divergent paths: Proprietary Generative AI Abstraction (exemplified by IBM watsonx Code Assistant for Z) and Open DevOps Ecosystem Orchestration (leveraging Zowe CLI, Endevor Web Services, and Jenkins). This analysis provides an architectural critique of both paradigms. This analysis we evaluate the cost-effectiveness, applicability, and system overhead of both methods using understanidng of a hybrid Zowe-Endevor-Jenkins pipeline. While GenAI offers rapid code translation, it introduces significant regression risk and compute overhead. Conversely, an open-source DevOps architecture modernizes the delivery framework without modifying stable business logic, probable better MIPS optimization and risk mitigation. Methodology & Architectural Scope The comparative evaluations, cost matrices, and pipeline workflows presented in this paper are the result of structural observations, technical analysis, and hands-on validation of isolated components (specifically, Endevor Web Services integration and Jenkins orchestration). This text is intended as an objective architectural blueprint for enterprise modernization. The complete, end-to-end 4-stage pipeline represents a proposed target state configuration based on current industry capabilities; multiple phases in this article are analyzed from a design perspective and have not been fully implemented or performance-tested as a unified system by the author. Introduction: The Modernization Dilemma The global financial and logistical ecosystem remains anchored to the mainframe. However, traditional application lifecycle management on z/OS—characterized by manual Interactive System Productivity Facility (ISPF) panels and linear promotion paths—creates severe operational bottlenecks. Enterprises attempting to modernize face a strategic binary choice: 1.Path A (Language Transformation): Deploy massive Large Language Models (LLMs) to abstract the mainframe by automatically translating legacy COBOL artifacts into cloud-native languages. 2.Path B (Pipeline Transformation): Expose existing z/OS systems to modern, distributed CI/CD orchestration layers, treating the mainframe as an agile backend node via open APIs. This paper analyzes these paths through the lens of architectural applicability and true total cost of ownership (TCO). Path A: The Proprietary GenAI Abstraction Paradigm IBM’s watsonx Code Assistant for Z represents the pinnacle of the GenAI abstraction strategy. Powered by localized Granite foundation models, this paradigm attempts to solve the skills gap by rewriting the software layer rather than the infrastructure pipeline. Architectural Mechanics The watsonx architecture ingests legacy enterprise code to perform automated application discovery, refactoring, and code generation [medium.com, medium.com]. It breaks down tightly coupled monolithic COBOL paragraphs into discrete services and outputs equivalent object-oriented Java code [medium.com, ibm.com]. System Overhead and Financial Realities While appealing to executive leadership, running large-scale LLMs for enterprise code refactoring demands intense computational infrastructure. Compute Spikes: Parsing millions of lines of COBOL triggers substantial CPU cycles. If executed on-platform, this risks inflating general-purpose MIPS (Millions of Instructions Per Second) consumption, directly impacting monthly software licensing costs via Tailored Fit Pricing models [royalcyber.com]. Hardware Dependendencies: Mitigating on-platform compute strain requires specialized hybrid-cloud routing or deep capital investment in on-premise hardware accelerators, such as the IBM Spyre Device [ibm.com]. The "Black Box" Testing Tax: The primary cost driver of the GenAI paradigm is not the tooling, but the subsequent validation. Automated translation introduces the risk of functional drift and code hallucination. Organizations must establish massive regression testing pipelines to ensure that auto-generated Java code perfectly mirrors the deterministic logic of the original COBOL asset. Path B: The Open DevOps Architecture The alternative path bypasses the risks of code mutation by standardizing the developer experience. Instead of translating the language, it exposes legacy environments to standard enterprise automation software. Core Case Study Component Breakdown This architecture is anchored by three independent layers that eliminate manual ISPF green-screen interventions entirely: 1.Endevor Web Services: Acts as the native mainframe lifecycle governance engine. By enabling REST API access to CA Endevor, the mainframe's inventory management (Add, Update, Move, Generate) is exposed as standard web endpoints. 2.Jenkins Automation Server: Serves as the distributed orchestrator. It manages the declarative workflows, security gating, and cross-platform synchronization pipelines. This helps to move the elements along the pipeline(different endevor stages) more efficiently following the current devops methodology. 3.Zowe CLI: The open-source command-line interface managed by the Linux Foundation's Open Mainframe Project [devops.com]. Zowe CLI allows external runners (like Jenkins agents) to communicate securely with z/OSMF or the Zowe API Mediation Layer, executing TSO/E commands, submitting JCL, and verifying dataset status using standard shell commands [docs.zowe.org, devops.com]. Proposed Implementation Arcitechture: The Hybrid Zowe-Endevor-Jenkins Pipeline To evaluate Path B against the GenAI model, we examine conceptual enterprise architecture. Stage 1: Development using GenAI Automated AI engines (like IBM watsonx Code Assistant for Z) connect directly to the z/OS host to discover, explain, and refactor code, injecting freshly generated source syntax directly into the development environment [medium.com, medium.com]. Advantage is rather than using expensive Code assistants, any open AI model can also be utilized in this scenario making it more cost effective. Based on published industry experience and practical observations, it reduces the code time and coding error by significant amount. Stage 2: Inventory Promotion via REST Endpoints Rather than a developer logging into a 3270 terminal emulator to execute a package promotion, the lifecycle begins with an automated trigger. The code undergoes modification and is pushed via a modern interface. An automated action invokes Endevor Web Services via an HTTP POST payload, executing a MOVE action to promote the software elements from the base development stage to the targeted lower-level QA/Testing stage in the Endevor map. Before implementing this workflow, organizations should verify whether their Broadcom Endevor environment supports REST API V1 or V2. (https://techdocs.broadcom.com/us/en/ca-mainframe-software/devops/ca-endevor-software-change-manager/18-1/using/using-the-rest-api/restful-api-v2.html ), the process needs to be changed to incorporate V2 API. There is an alternative to it is to use Endevor Bridge for Git where the codebase is synced with Git and mainframe repo is created. So on place of coding in mainframe the coding is done via changing in the local repo then pushing to git using pull request. That will eventually sync to endevor through the Git and endevor Bridge. Upon successful element promotion, the Endevor Webhook Server broadcasts a status payload to the enterprise automation network. Stage 3: Jenkins Orchestration & Quality Gates The incoming webhook triggers a declarative Jenkins pipeline. Jenkins initializes an isolated workspace container configured with Node.js and the Zowe CLI core framework along with the Endevor plug-in. One possible implementation sequence could include, before any mainframe deployment activities commence, Jenkins routes the source code through modern enterprise quality gates (e.g., SonarQube) to check syntax compliance and security vulnerabilities. Stage 4: Recommended Target Deployment & Execution Framework Once quality gates clear, Jenkins switches from analysis to active infrastructure execution. It issues structured Zowe CLI scripts to perform runtime deployment tasks across target Logical Partitions (LPARs): Automated Package Shipping: Jenkins fires zowe endevor ship package to distribute compiled binaries to remote testing LPARs. Database Binding: The pipeline automatically identifies modified Database Request Modules (DBRMs) and executes zowe zos-files submit job to trigger remote JCL for required DB2 binds. Runtime Environments: Finally, Zowe CLI sends direct system commands to recycle lower-level CICS regions or clear IMS transaction queues, instantly rendering the newly promoted code active for automated functional test suites. The Critical MIPS and Cost Divide The defining architectural difference between these paths lies in compute economics. Path A introduces a brand new computational workload (AI inference) to track down and translate legacy code structures [ibm.com]. Path B, through the Zowe API Mediation Layer, is built explicitly for system efficiency. By running the orchestration engine on external Jenkins nodes and transmitting lightweight JSON payloads to the mainframe, administrative tasks are streamlined. Furthermore, interactions with z/OSMF via Zowe CLI can be offloaded directly to System z Integrated Information Processors (zIIP). This prevents spikes on general-purpose central processors, shielding the organization from surging software usage fees [royalcyber.com]. Synthesis: The Real Solution at the Crossroads Choosing between Path A and Path B is a false dichotomy. The core flaw of the proprietary GenAI abstraction approach is its operational isolation. If an advanced LLM generates flawless Java code from a legacy COBOL source, but the organization still relies on manual, legacy deployment frameworks to push that code into production, the modernization effort has failed. Therefore, an Open DevOps ecosystem can be proven more effective for successful mainframe modernization. Even if an enterprise chooses to leverage IBM watsonx for selective code refactoring, it requires the structural foundation of a Zowe, Endevor Web Services, and Jenkins pipeline to ingest, validate, bind, and deploy those assets automatically. Conclusion Enterprise infrastructure modernization must favor deterministic execution, predictability, and rigorous cost optimization. While Generative AI tools like watsonx offer intriguing possibilities for future code refactoring, they cannot replace a modern delivery pipeline. Also the years of complexity/business logic incorporated in mainframe code need human in loop to review the implemented and modernized code in the distributed services. Also that incur additional cost for running data certification, regression and parallel running 2 systems in both mainframe and new technologies until the new system stabilies or proven correct. The integration of Zowe CLI with Endevor Web Services and Jenkins proves that legacy mainframe ecosystems do not need to be completely rewritten to achieve modern agility. By upgrading the delivery engine instead of compromising the application logic, enterprises can achieve true DevOps velocity, safeguard operational budgets, and successfully navigate the modernization crossroads. Disclaimer: The views expressed are solely those of the author. This article reflects my personal technical opinions and analysis based on publicly available technologies and general enterprise engineering practices. It is not endorsed by or representative of any employer. References used : https://planetmainframe.com/2026/01/agentic-ai-on-the-mainframe-greater-delivery-flexibility-and-more/ https://www.zowe.org/ https://www.ibm.com/new/announcements/ibm-watsonx-code-assistant-for-z-adds-ai-code-generation-and-assembler-support https://www.royalcyber.com/blogs/mainframe/mainframe-cost-optimization-mips-reduction/ https://croz1.substack.com/p/marvelous-mainframe-playground https://techdocs.broadcom.com/us/en/ca-mainframe-software/devops/endevor-bridge-for-git/2-0/using/access-and-clone-an-initialized-git-endevor-mapping.html