Programming Languages in the Age of "AI" Agents (alexn.org)

As “AI” agents increasingly generate code, the author argues programming language choice still matters — perhaps more than ever. Popular languages like Python benefit from large training corpora and yield usable agent-generated scripts, but expressive static type systems (Scala, Haskell, Rust) give agents a much faster, more reliable feedback loop: compile errors and LSP integration (e.g., Metals + Copilot) let agents iterate until code compiles, catching whole classes of bugs that unit tests might miss. That reduces hallucination risk and makes the agent workflow practical for real development, not just toy scripts. However, reliance on agents creates new risks: incomplete tests, fragile documentation of intent, and “comprehension debt” when no human understands the evolving design. Agent dialogs have limited context windows and can suffer context poisoning, so generated code plus generated tests aren’t a substitute for clear, high-level source that captures design invariants. The takeaway for the AI/ML community: favor languages and practices that make intent explicit (strong typing, higher-level abstractions, functional/equational reasoning) and invest in human-reviewable code, tests, and architecture. These measures preserve maintainability and ensure agents augment — rather than erode — long-term software quality.