LLM-Powered Code Porting Shocks Developers: 70,000 Lines in 3 Days – Experts Rethink Legacy Migration Strategy

In a breakthrough demonstration at a private industry retreat, a development team created a behavioral clone of the GNU COBOL compiler in Rust—generating 70,000 lines of code in just three days using large language models (LLMs). The rapid port underscores the accelerating ability of AI to lift and shift legacy systems, sparking urgent debate among software engineers about the future of migration strategies.

“This is another sign of the ability of LLMs to do a good job of porting existing code to a new platform,” said a source familiar with the project, who spoke under the condition of anonymity. The success, they noted, hinged on the availability of strong regression tests—though the quality of GNU COBOL’s test suite remains unclear. “If you have access to an existing implementation, you can build a test suite from it,” the source added.

The achievement was one of several revelations at The Orchard Retreat, hosted by Mechanical Orchard. The event convened industry leaders to discuss the profession’s future amid the rise of agentic programming. All discussions were held under the Chatham House Rule, so attendees’ identities remain confidential.

Key Developments

‘Interrogatory LLM’ System Presented

Another attendee described a novel technique for verifying large specification documents: an LLM “interviewing” a human expert to confirm the correctness of each requirement. This approach, dubbed the Interrogatory LLM, aims to reduce human error when reviewing complex specs.

“The LLM asks questions to ensure no detail is overlooked,” the source said. The method could improve accuracy in sectors like finance and defense where spec fidelity is critical.

‘Scar Tissue’ of Change-Control Boards

One consultant shared a practical insight: upon arriving at a new organization, they first read the guidelines of the change-control board. “This is the scar tissue of what’s gone wrong in the past,” they noted. The advice resonated with many attendees, who agreed that understanding an organization’s history is essential to making effective changes.

The Great ‘Lift and Shift’ Debate

A long-standing point of contention among legacy modernization experts—the wisdom of “lift and shift” (porting a system to a new platform while preserving all features)—was revisited in light of LLM capabilities. Traditionally, many consultants have warned against it, arguing that old systems often contain 50% unused features (according to a 2014 Standish Group report) and outdated business processes.

“Replacing these features is a waste,” said one specialist. “Instead, take a step back and understand what users actually need.” However, that view was developed before LLMs could cheaply automate porting. “Now, lifting and shifting should be always the first step in a legacy migration,” said an attendee who works extensively in the field. “The cost is no longer formidable, and a better platform makes further evolution much cheaper. Just don’t stop there.”

Background

The Orchard Retreat, held the week of May 14 in an undisclosed location, gathered roughly 30 professionals from software development, finance, and consulting. The event focused on how agentic programming—using AI agents to write and manage code—is reshaping development practices. Participants included engineers from major financial institutions grappling with complex legacy environments, regulatory controls, and high operational risk.

Discussions were strictly off the record, but multiple sources corroborated the key findings reported here. The retreat’s hosts, Mechanical Orchard, did not respond to requests for comment.

What This Means

The ability to port a complete compiler in three days signals a paradigm shift in legacy modernization. Companies that previously avoided lift and shift due to high costs may now adopt it as an initial step, using LLMs to translate COBOL, FORTRAN, or other ancient languages into modern counterparts like Rust or Go. This could dramatically reduce migration timelines and budgets, while freeing engineering teams to focus on feature redesign.

However, experts caution that regression test coverage remains critical. Without robust tests, automated porting can introduce subtle bugs. The ‘interrogatory LLM’ technique offers a way to validate specifications but is not yet widely adopted. “The industry is in a race between AI capabilities and the need for rigorous verification,” said one financial sector attendee. “We should be both excited and careful.”

In the long term, the retreat’s insights suggest that organizations will need to combine AI-driven porting with human oversight to avoid repeating past mistakes—and to truly modernize, not just move code around. As one participant put it: “We now have the tools to lift and shift cheaply. The real challenge is having the courage to stop and ask, ‘What do users actually need?’”