Martin Fowler shares notes from a Chatham House-rule retreat on the future of software development in the age of agentic programming. Key observations include: a team building a GNU Cobol behavioral clone in Rust in 3 days using LLMs; using LLMs to interview human experts for spec verification; a reassessment of 'lift and shift' legacy migration now that LLM-assisted porting is cheap; financial industry challenges around multi-jurisdiction regulatory complexity and whether per-jurisdiction systems could be maintained with LLM consistency checks; pair programming as a vehicle for skills transfer in the agentic era; and the idea of a 'Chaos Monkey for AI' to test hallucination detection. Fowler also links to arguments that developers overuse agents at runtime when LLM-as-function is more predictable, and that skill files are overused when better code architecture would suffice. He reflects on Kyle Kingsbury's pessimistic long-form essay on AI's societal trajectory, sharing his own 'hoper and doomer' stance.
Sort: