Why AI-Powered Code Migration Doesn't Scale Without Agent Systems

Why AI-Powered Code Migration Doesn't Scale Without Agent Systems

AI is already remarkably good at migrating individual code components. For large systems, however, a single prompt is not enough. AI-powered code migration only becomes scalable when analysis, prioritization, code generation, testing, quality assurance, and governance work together in a controlled process.

For companies with complex, evolving IT landscapes, this is the decisive factor. Modernization efforts fail due to unclear dependencies, a lack of transparency, undefined priorities, inadequate testing, and the question of how changes are documented in a traceable manner.

Management Summary

AI-assisted code migration can significantly accelerate modernization projects if it is not viewed merely as prompting and code translation.

Individual LLMs are well-suited for clearly defined tasks. However, in large system landscapes, without additional control, there is a lack of oversight regarding dependencies, priorities, testing, and approvals.

Agent systems help bridge this gap. They handle specialized tasks such as analysis, classification, code generation, test validation, error handling, and documentation.

The greatest benefit lies in a migration process that is more predictable, verifiable, and easier to control.

Regulated industries in particular—such as banking, insurance, pharmaceutical companies, and public organizations—require traceability and auditable systems. Every technical change must be explainable, verifiable, and documentable.

Why Traditional Prompt Engineering Isn't Enough for Legacy Migrations

Simple prompts work well when the task is clearly defined: understanding a module, translating a class, creating a test file, or applying a known pattern.

In a large-scale migration, however, the task is rarely isolated. Programs access shared data structures and a wide variety of heterogeneous data sources. The business logic is distributed across scripts, workflows, data flows, and manual processes. On top of that, documentation is typically incomplete or outdated.

Code, therefore, cannot simply be translated from A to B. First, you must understand:

• Which systems are business-critical?
• Which applications are functionally or technically dependent on one another?
• Which components can be migrated automatically?
• Where is human judgment required?
• What tests confirm that the business logic is preserved?
• What changes must be documented for audit, governance, and change management?

A single LLM can assist with these tasks. However, it cannot reliably manage them, as outputs may vary from user to user even with the same input.

Migration is not a traditional software engineering project

In large-scale modernization projects, pure code translation is only part of the work. The greater effort comes from analysis, classification, prioritization, quality assurance, business coordination, and risk management.

This is especially true in regulated industries. In these sectors, it’s not enough for the result to be correct; the process leading to that result must also be traceable. Companies must be able to explain what was changed, why it was changed, which risks were assessed, and who approved the decision.

A purely prompt-based approach reaches its limits here.

Typical vulnerabilities include:

What Sets Agent Systems Apart

Agent systems do not view AI as a tool for individual tasks. Rather, they are part of a controlled workflow in which different agents take on specialized tasks within the migration process.

One agent can analyze source code. Another classifies risks and dependencies. Yet another generates target code. Additional agents create tests, verify results, document decisions, or flag cases for human approval.

Of course, agents do not fully automate a migration. The advantage lies in the fact that they structure recurring tasks, make results verifiable, and involve human experts in a more targeted manner.

For companies, this leads to a different approach to modernizing complex system landscapes:

• Migrations become more predictable because systems and artifacts are systematically classified.
• Risks become apparent earlier because dependencies and testing gaps are documented.
• Quality becomes easier to verify because testing and validation are part of the process.
• Decisions and results are recorded in a traceable manner, making the entire project more controllable and transparent.
• Technical experts are involved only where human judgment is truly necessary.

What Real-World Examples Show

Current real-world examples show that LLM-supported migrations can work in larger software codebases when backed by automation, testing, and human oversight.

Airbnb migrated approximately 3,500 React test files from Enzyme to the React Testing Library. The manual effort was originally estimated at about 1.5 years. With LLM-assisted automation, the migration was completed in six weeks.

In a research report, Google describes 39 internal code migrations over a twelve-month period. During this time, 595 code changes with 93,574 edits were submitted. A significant portion of the changes was generated by LLMs and subsequently reviewed in an automated workflow.

Slack describes the transition from more than 15,000 Enzyme test cases to React Testing Library. Here, too, it was not a single prompt that was decisive, but rather the combination of traditional code analysis, automation, LLM support, and human oversight.

These examples are important, but they should be put into proper context. They demonstrate that LLMs can be effective in large-scale migrations when embedded within a controlled process.

The Actual Business Benefit

The added value of AI-supported code migration thus lies in better control over the entire modernization project.

For companies, this means:

• less manual routine work for recurring transformation tasks
• better predictability through systematic analysis and classification
• lower implementation risk through automated testing and validation
• greater transparency for governance, auditing, and change management
• more targeted use of internal business and IT resources

This can be decisive for the success of a modernization program, especially in large and complex legacy environments. Without structure, AI quickly becomes a tool with no added value in an already complex project.

What Companies Should Consider Before Launching

The most important question is whether the conditions for a scalable migration are in place.

Five points should be clarified before getting started:

1. Which systems and applications are business-critical?
2. What dependencies need to be understood before the migration?
3. What is the binding target architecture?
4. What testing strategy demonstrates functional equivalence?
5. Which decisions must be documented and approved?

A good place to start is with an assessment before beginning the actual migration. An assessment analyzes the existing application landscape, the complexity of the source code, dependencies, and business-critical processes to establish a solid foundation for modernization and migration.

The Role of HMS

HMS supports companies in planning complex migration projects in a structured manner, implementing them technically, and ensuring their functional integrity. The focus is on the entire process, from analysis to the validated target environment.

In regulated industries in particular, an approach is needed that combines technical efficiency with traceability. This includes a clear assessment of the initial situation, a robust target vision, appropriate automation, quality assurance, and documented decisions.

Conclusion: It’s not the prompt that scales the migration, but the process

AI-powered code migration works particularly well when it is not viewed as a standalone task. A language model can make developers more productive. A controlled process with specialized agent systems can make a migration program more predictable, verifiable, and scalable.

For companies with large legacy landscapes, this is the decisive difference.

The next sensible step, therefore, is not a large-scale AI rollout, but a structured assessment aimed at establishing a clear baseline, representative artifacts, a defined target architecture, and measurable quality criteria—upon which a structured migration can be built, true to the motto:

Better to do it right once than three times quickly.

Are you facing a migration and want to get off to a solid start?

You can find more information about our approach in our Code2X white paper.

Download

Sources and Further Reading

• Airbnb Engineering: Large-Scale Test Framework Migration Using LLMs (2024)
• Platform Engineering: Are AI Agents Ready for Large-Scale Migrations? (Dec. 2025)
• Google Research: Migrating Code at Scale with LLMs at Google (2025)
• Abto Software: AI Code Migration: LLMs Explored (2026)