Beyond the Code: Insights from a Software Development Retreat on Agentic Programming
The Rise of Agentic Programming: A Gathering of Minds
Last week, a small group of software professionals convened for a day-long retreat to explore how agentic programming is reshaping the craft. Held under the Chatham House Rule, the conversation was candid and attribution-free—though participants who recognize their contributions are welcome to claim them. The discussions yielded a handful of sharp observations that deserve a wider audience.
LLM-Driven Porting: The GNU Cobol Example
One team shared a striking experiment: they used an LLM to create a behavioral clone of the GNU Cobol compiler entirely in Rust. The resulting codebase was 70,000 lines of Rust, built in just three days. This demonstrates how effectively LLMs can port existing code to new platforms—especially when backed by strong regression tests. (The quality of GNU Cobol's test suite wasn't disclosed, but the principle holds: if you have a solid implementation, you can even generate a test suite from it.)
Interrogatory LLMs: A New Way to Validate Specifications
Large specification documents are notoriously difficult for humans to review thoroughly. One attendee proposed a clever workaround: have the LLM interview a domain expert. Instead of asking the expert to read every line, the system poses questions to verify correctness—a method they called Interrogatory LLM. This turns passive review into an active, targeted dialogue.
The Scar Tissue of Change Control
Not every insight was AI-focused. One consultant mentioned that their first step when working with a new organization is to read the change-control board's guidelines. They described these rules as the scar tissue of past failures—a living record of what went wrong. It's a powerful reminder that to understand why a process is the way it is, you must first learn the history that shaped it.
Lift and Shift Reconsidered
For years, my colleagues in legacy modernization dismissed lift and shift—porting a system to a new platform while preserving feature parity—as a missed opportunity. After all, legacy systems often accumulate bloat; the Standish Group's 2014 report found that up to 50% of features go unused. Why waste effort replicating them? The better approach, we argued, was to step back, identify real user needs, and prioritize against business outcomes.
But LLMs have changed the calculus. An attendee deeply involved in legacy work argued that lift and shift should now always be the first step. The cost of porting has plummeted, and once the code runs on a modern platform, further evolution becomes far cheaper and safer. The key is not to stop there—use the improved environment to gradually refactor and slim down the system.
Legacy Systems in Finance: A High-Stakes Environment
Several attendees hailed from the financial sector, where legacy systems intertwine with strict regulatory controls and immense risk. A single software error can lead to millions in losses. For them, the LLM's ability to port code accurately is a lifeline—but it also raises the stakes for validation and testing. The conversations around agentic programming in such environments were especially intense, highlighting the tension between innovation and reliability.
These fragments from the retreat offer a glimpse into how the software industry is grappling with the agentic era. From accelerated porting to smarter specification reviews, the landscape is shifting—and the old rules no longer apply.