LLM-Powered Code Migration Achieves 70K Lines in 72 Hours
In a stunning demonstration of artificial intelligence capabilities, a team of developers has successfully ported the GNU COBOL compiler to Rust, creating a 70,000-line behavioral clone in just three days. The achievement marks a seismic shift in legacy system modernization, as large language models (LLMs) now enable what once took months or years.
The work was revealed during an off-the-record industry retreat focused on agentic programming. Attendees spoke under Chatham House Rule, meaning their identities cannot be disclosed, but multiple sources confirmed the feat to this publication.
Background: The Retreat and the Challenge
The retreat gathered dozens of software professionals from finance, consulting, and technology firms. The topic: how AI agents are reshaping software development, particularly the painful process of porting decades-old code.
Legacy systems, especially in banking and government, often run on languages like COBOL that are difficult to maintain. The traditional approach—"lift and shift"—was widely criticized as a missed opportunity, leaving bloated, unused features untouched.
Key Findings and Expert Quotes
Behavioral Clone Success
"One group built a complete behavioral clone of GNU COBOL in Rust using an LLM," an attendee with direct knowledge said. "The result is 70,000 lines of Rust, assembled in three days. That's a 24x speedup over manual porting."
The source emphasized the value of robust regression tests. "If you have a good test suite, LLMs can validate correctness automatically. If not, you can generate tests from the original implementation."
Interrogatory AI for Specifications
Another attendee shared a novel use of LLMs: treating them as interrogators for specifications. "Rather than humans struggling through a 500-page spec, the AI interviews an expert, asking targeted questions to verify each requirement," the source explained. This approach could drastically reduce review time.
The Scar Tissue of Change Control
A consultant noted that their first step when entering an organization is reading the change-control board's guidelines. "That documentation is the scar tissue of every past mistake," they said. "Understanding why things are the way they are is critical before proposing changes."
Reevaluating 'Lift and Shift'
The retreat produced a revisionist view on legacy migration. Previously, experts advocated against simply porting code without rethinking features. A 2014 Standish Group study found that 50% of features in old systems are unused.
"But LLMs change the economics," a legacy modernization specialist argued. "Lift and shift should now always be step one. The cost is no longer prohibitive. A modern platform makes subsequent evolution orders of magnitude cheaper." The source cautioned: "Just don't stop there—use the new environment to trim bloat and align with business metrics."
What This Means for the Industry
Financial institutions, burdened by regulatory constraints and high risk of monetary loss, stand to benefit most. "The barrier to moving off COBOL and onto maintainable stacks just collapsed," said a senior architect from a major bank.
The ability to port code in days means organizations can modernize incrementally. However, experts warn that without accompanying process improvements and feature reduction, companies risk simply replicating old problems on new platforms.
The takeaway: LLMs are not a silver bullet, but they remove the biggest hurdle—cost and time—to starting a migration. The industry now has a powerful new tool in the fight against technical debt.
— Reporting contributed by multiple retreat attendees under Chatham House Rule.