Breaking News: LLMs Port GNU Cobol to Rust in 72 Hours
In a startling demonstration of AI's capability, a team of developers has created a behavioral clone of the GNU Cobol compiler using Rust—in just three days. The resulting codebase amounts to 70,000 lines, underscoring the accelerating role of large language models (LLMs) in legacy system modernization.
“This is a clear signal that LLMs can efficiently port existing code to new platforms,” said an attendee who witnessed the project. “But the quality of regression tests makes or breaks the outcome.” The breakthrough was shared during a recent retreat focused on the future of software development amid the rise of agentic programming.
The Retreat: A Closed-Door Summit on Software's Future
Held under the Chatham House Rule, the event at The Orchard Retreat, hosted by Mechanical Orchard, gathered dozens of developers, architects, and industry veterans. Discussions ranged from code portability to risk management in legacy environments. Because of the rule, individual comments cannot be attributed, but several key ideas emerged.
Interrogatory LLM: A New Way to Verify Specs
One participant proposed using an LLM to interview a human expert, asking clarifying questions about large specification documents. The technique, dubbed “Interrogatory LLM,” aims to surface errors and gaps that a human reviewer might miss. “It turns the AI into an active audit tool,” the attendee explained.
Change-Control Boards as ‘Scar Tissue’
Another insight focused on organizational history. “The first thing I do when consulting is read the change-control board guidelines,” said one veteran. “That’s the scar tissue of everything that went wrong before.” The comment highlights how institutional memory is often embedded in policies, not just code.
The Great Lift-and-Shift Debate
Before LLMs became widely capable, many experts dismissed “lift and shift” as a wasted opportunity. The prevailing view held that porting legacy systems to new platforms without rethinking features ignored decades of bloat. According to a 2014 Standish Group report, up to 50% of software features are rarely or never used.
But LLMs have changed the calculus. “Now, lift and shift should always be the first step,” argued one attendee specializing in legacy migrations. “The cost is no longer prohibitive, and a modern platform makes further evolution vastly cheaper—just don’t stop there.”
Background
The retreat was convened to address the intersection of agentic programming and software engineering. Agentic programming refers to AI systems that can autonomously plan and execute coding tasks. Participants represented industries from finance to cloud services, each grappling with aging codebases and regulatory pressures.
Financial sector attendees highlighted the tension between innovation and compliance. “Complex legacy systems, strict regulatory controls, and high risk—those are our daily reality,” one said. The discussion underscored that while AI can accelerate migration, governance must keep pace.
What This Means
The ability to port a compiler in three days signals a paradigm shift. For organizations stuck on mainframes or obsolete languages like Cobol, LLMs offer a fast, cost-effective path to modern infrastructure. However, success hinges on robust test suites and a willingness to evolve after migration.
Experts urge caution: “Lift and shift is not a final destination,” warned another attendee. “It’s a stepping stone. Use the new platform to prune unused features and align with current business needs.” The retreat concluded with a call for more collaboration between AI tooling developers and domain experts.
As one participant summarized: “AI won’t replace developers—but it will radically change what they spend their time on.”