AI Code Porting Breakthrough: COBOL Compiler Rebuilt in Rust in Record 3 Days
Software Revolution: LLM-Powered Migration Sets New Speed Record
A team of developers has achieved a stunning feat in software engineering, using large language models (LLMs) to create a behavioral clone of the GNU COBOL compiler in Rust—a task that produced 70,000 lines of new code in just three days. The breakthrough, revealed at a confidential industry retreat, underscores the transformative potential of LLMs in legacy system modernization.

“This is another sign of the ability of LLMs to do a good job of porting existing code to a new platform,” said one attendee familiar with the project. “But it also highlights the critical importance of having a robust regression test suite—although we don’t know how good GNU COBOL’s tests are.” The rapid porting effort suggests that even complex, decades-old compilers can be migrated with minimal manual intervention.
‘Interrogatory LLM’ Emerges as Verification Tool
In another development, a participant proposed using an LLM to interview human experts to verify the correctness of large specification documents. “The idea is to have the LLM ask targeted questions of a domain expert, essentially cross-examining the specification to confirm its accuracy,” the source explained. This approach—dubbed “Interrogatory LLM”—could reduce human review fatigue and catch errors in complex documentation.
Change-Control Guidelines: The Scar Tissue of Failure
Beyond AI-specific insights, one attendee shared a universal truth about organizational change. “The first thing I do when consulting with an organization is to read the guidelines for their change-control board,” the expert said. “That’s the scar tissue of everything that’s gone wrong in the past. Understanding the history of how a system got to its current state is essential to fixing it.”
Legacy Modernization: Rethinking the ‘Lift and Shift’ Strategy
The retreat also sparked a fresh debate on “lift and shift”—the practice of moving a legacy system to a new platform while preserving feature parity. Previously dismissed by many modernization specialists as a missed opportunity, the approach is now gaining new respect thanks to LLMs. “We used to say lift and shift was a waste because legacy systems are bloated with unused features,” explained another participant. “But now, with LLMs making porting so cheap, one attendee argued it should always be the first step. Just don’t stop there—use the new, better environment to evolve.”
The shift in perspective is significant for industries like finance, where multiple attendees operate under strict regulatory constraints. “Complex legacy environments coupled with regulatory controls and significant risk are our daily reality,” said a financial sector expert. “If LLMs can safely accelerate the first migration step, it changes the entire modernization calculus.”
Background
The revelations emerged from “The Orchard Retreat,” a private gathering hosted by Mechanical Orchard that convened software professionals to discuss the profession’s future amid the rise of agentic programming. The event operated under the Chatham House Rule, meaning comments can be reported but not attributed to specific individuals. Participants included developers, consultants, and technology leaders from the financial industry and beyond.
The retreat took place less than two weeks ago, and this is the first public summary of its key insights. The timing is crucial as organizations worldwide scramble to understand how LLMs will reshape software engineering, particularly for maintaining and upgrading critical legacy systems.
What This Means
The ability to port a full compiler in three days demonstrates that LLMs have reached a threshold where large-scale code migration is not only feasible but dramatically faster and cheaper than human-driven rewrites. This could accelerate the modernization of the estimated 220 billion lines of COBOL code still running in banking, government, and insurance systems.
However, the reliance on regression tests remains a bottleneck. Organizations that lack comprehensive test suites may need to invest in automated test generation before reaping the benefits of LLM porting. The “Interrogatory LLM” concept offers a path forward for validating specifications, but it is still early in its development.
For legacy modernization consultants, the new advice is clear: reconsider “lift and shift” as a viable first step. As one attendee summarized, “Don’t stop at porting—that’s just the beginning. But if the cost barrier is gone, do it first to create a better foundation for everything else.” Financial institutions, in particular, must balance the speed of LLM-driven migration with the strict compliance and audit requirements that govern their operations.
Overall, the software engineering profession is at an inflection point. The tools now exist to tackle historically intractable problems of code migration and specification verification—but human expertise, especially in regulatory contexts, remains irreplaceable.
Related Articles
- VS Code Python Extension Update: Enhanced Code Navigation and Faster Indexing (March 2026)
- How to Contribute to the Python Insider Blog with Git and Markdown
- Scaling Multi-Agent AI Systems: Overcoming Coordination Challenges in Large-Scale Deployments
- Why Domain Expertise Remains Critical in the Age of AI-Assisted Development
- Kubernetes v1.36 Delivers Declarative Validation: What It Means and How It Works
- 5 Must-Know Governance Features for .NET AI Agents with MCP Tool Calls
- Google's AI Crisis Deepens as Developer Conference Looms Amid Coding Tool Struggles
- Cursor vs Windsurf: A Comprehensive Python Developer's Guide to Choosing the Right AI Code Editor