Agentic Programming and Legacy Systems: Insights from a Developer Retreat

Last week, a gathering of software developers convened to discuss the future of the profession as agentic programming emerges. Held under the Chatham House Rule, the retreat offered a candid exchange of ideas on how large language models (LLMs) are reshaping everything from porting legacy code to verifying specifications. Here are the key takeaways, synthesised from the discussions.

Behavioral Clone of GNU COBOL in Rust

One team reported creating a behavioral clone of the GNU COBOL compiler written entirely in Rust. The result: 70,000 lines of Rust code produced in just three days. This feat underscores the growing ability of LLMs to port existing software to new platforms efficiently. However, such rapid translation relies on robust regression tests to ensure correctness. The original GNU COBOL compiler likely has its own test suite, but the exercise also suggests that if a test suite is missing, it can be generated from an existing implementation. This approach dramatically reduces the cost and risk of migrating legacy systems.

Interrogatory LLM: Validating Specifications Through Dialogue

Large specification documents are notoriously difficult for humans to review thoroughly. One attendee proposed an innovative solution: using an LLM to interview a domain expert. The AI would question the human to verify each part of the specification, acting as a form of “interrogatory LLM.” This method turns a static review into a dynamic conversation, catching inconsistencies and ambiguities that might otherwise slip through. It’s a practical application of AI that augments human expertise rather than replacing it.

Change-Control Boards as Organizational Memory

Not directly related to AI, but a valuable insight: one consultant mentioned that the first thing they do when working with an organization is to read the guidelines of its change-control board. These documents represent the “scar tissue” of past failures—every rule exists because something went wrong before. Understanding this history is often the key to understanding why systems are the way they are. For developers tackling legacy environments, reviewing change-control board records provides essential context that can prevent repeating old mistakes.

Rethinking Lift-and-Shift in the AI Era

For years, proponents of legacy modernization have criticized “lift and shift”—porting a system to a new platform while preserving feature parity. The argument against it has been strong: legacy systems often contain bloated codebases where up to 50% of features are unused (according to a 2014 Standish Group report). Simply moving everything wastes resources; instead, teams should analyze what users actually need and prioritize those features against business outcomes.

But the conversation at the retreat introduced a new perspective driven by LLMs. With the ability to port code quickly and cheaply, one attendee argued that lift and shift should now be the first step in any legacy migration. The cost has dropped so dramatically that it makes sense to move the entire system to a modern platform first, then evolve it incrementally. The key is not to stop there—use the new environment’s flexibility to refactor and remove unnecessary features.

Challenges in the Financial Sector

Several participants came from the financial industry, where legacy systems are deeply intertwined with regulatory controls and significant risk. Mistakes in software handling money can have severe consequences. The combination of complex legacy code, strict compliance requirements, and high stakes makes financial institutions particularly cautious. Yet, as LLMs prove their ability to port and document code accurately, even these risk-averse organizations may begin to adopt AI-assisted modernization—provided the testing and validation processes are rigorous enough.

Looking Ahead

The retreat highlighted a shift in mindset: instead of fearing AI as a disruptor, developers are finding practical ways to harness it for solving long-standing problems. From behavioral clones and interrogatory reviews to rethinking lift-and-shift, the profession is learning to work with AI as a powerful collaborator. The key is to maintain a focus on quality, testing, and understanding the historical context of each system.

These insights were shared under the Chatham House Rule; if you recognize your contribution and wish to be attributed, please reach out.