Fragments: May 14

2026-05-14 · Source: Martin Fowler · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Advanced, long

Summary

The "Fragments: May 14" brief compiles observations from a Mechanical Orchard retreat on agentic programming's impact on software development. Key discussions included an LLM-powered behavioral clone of a GNU Cobol compiler in Rust, completed in 3 days with 70K lines of code, highlighting LLMs' code porting ability. The concept of "Interrogatory LLMs" for human-expert specification verification was explored. Attendees re-evaluated "Lift and Shift" legacy migrations, suggesting it as a first step due to reduced LLM-driven costs. The potential for LLMs to simplify multi-jurisdictional financial systems and automate tedious data transformation was noted. Concerns arose regarding teaching judgment to junior developers in agentic contexts, with pair programming suggested. Critiques of agentic programming emphasized using LLMs as predictable functions over autonomous agents and prioritizing architectural solutions over "skills" for reliability. The article also touches on the non-deterministic nature of AI and ethical considerations, referencing Kyle Kingsbury's "The Future of Everything is Lies, I Guess."

Key takeaway

For software architects and MLOps engineers designing new systems or modernizing legacy ones, recognize that LLMs fundamentally alter cost-benefit analyses for tasks like code porting and data transformation. Prioritize integrating LLMs as predictable functions within defined workflows rather than relying on autonomous agents, which introduce debugging complexity. Emphasize human oversight and learning from AI-generated outputs, especially in critical review processes, to ensure long-term skill development and system reliability. Consider "lift and shift" as a viable initial step for legacy modernization, leveraging LLMs to reduce costs and enable easier future evolution.

Key insights

LLMs are transforming software development by enabling rapid code porting and data transformation, but require careful integration to maintain predictability and foster human learning.

Principles

• LLMs excel at code porting and data transformation.
• Human learning from AI choices is crucial for skill growth.
• Prioritize architectural solutions over LLM "skills."

Method

Interrogatory LLM: LLM interviews human expert to verify specification correctness. Legacy migration: Lift and shift to new platform first, then evolve.

In practice

• Use LLMs for code porting and data transformation tasks.
• Employ pair programming to mentor junior agentic developers.
• Design LLM applications as predictable function calls, not autonomous agents.

Topics

• Agentic Programming
• LLM Applications
• Code Modernization
• Software Architecture
• Legacy Systems
• AI Ethics
• Pair Programming

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Software Engineer, MLOps Engineer

Related on AIssential

• Trump’s AI Order Won’t Stymie U.S. Competition with China — A future Trump AI order is predicted to not hinder U.S. AI competition with China.
• EU Unveils Sweeping Tech Sovereignty Push, Balancing Autonomy with Openness — Europe's tech sovereignty strategy balances reducing foreign dependency with fostering domestic capacity and open markets.
• Personalization is the #1 Driver of Meaningful Employee Recognition, Snappy Workforce Study Finds — Personalized recognition, particularly through gifting, is the top driver for meaningful employee appreciation and engagement.
• Micro-Dyn, a leader in healthcare claims pricing software, announces its acquisition of CMSPricer. — Micro-Dyn's acquisition of CMSPricer expands its healthcare claims pricing and contract management platform with new technology and executive expertise.
• ThroughPut.AI Now Available on SBIR Aisle of CDAO’s Tradewinds Solutions Marketplace — AI-driven materiel flow optimization significantly improves military equipment readiness by preempting shortages and streamlining maintenance.
• Asset Panda Introduces “Build with Excel” to Streamline AI-Powered Account Setup in Asset Panda Pro — Asset Panda's "Build with Excel" uses AI to convert existing spreadsheet data into a functional asset tracking system, simplifying setup.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Martin Fowler.