The bitter lesson for software

A provocative essay titled 'The bitter lesson for software,' posted by authors zef, rohuang, and kaivu on the AI forum LessWrong, has gone viral by framing the rise of AI agents as an existential shift for traditional software development. The core argument is that software historically succeeded by encoding useful human actions into rigid, deterministic information flows—think ERP systems managing procurement or Git managing code changes. However, AI agents represent a new class of 'agent software' that encodes information flows with far greater flexibility. These agents, built on models like GPT-4 or Claude 3, can execute on open-ended commands and work with the natural complexity of real-world tasks without first needing to compress that complexity into predefined data structures. They combine system-specific knowledge with generalized pre-training, all while inheriting software's benefits like rerunnability and scalability.

The essay posits that this marks the arrival of the 'bitter lesson'—a concept from machine learning where general, scalable methods consistently outperform systems built with hand-crafted human knowledge—to the entire software stack. The space for building classical, deterministic software is shrinking. Instead, future software will increasingly be 'agent software': an infrastructural backbone of code that orchestrates various AI agents in flexible, sometimes newly synthesized, workflows. Low-level, repeatable components will remain in code, but higher-level logic and decision-making will be handled by agents. Concrete examples include using agents for adaptive, full-coverage feature testing in CI/CD pipelines instead of static test suites, or for flexible research-grade web scraping and analysis that can be repurposed across different studies. The conclusion is that agent software will 'eat' the well-defined world of pre-AI software, spitting it out 'softer' and more adaptable, fundamentally altering how we solve problems with technology.