Compose tasks, not implementations (2025) (thesephist.com)

A new perspective on designing complex AI systems emphasizes composing tasks rather than focusing on implementation details. This approach advocates for defining abstractions around specific tasks and subtasks a system should perform, allowing for more robust and maintainable AI architectures. By defining a system as a composition of tasks—akin to composing functions in programming—the designer can localize changes and failures to particular tasks, thus making systems easier to understand and evolve over time. The significance of this method lies in its potential to mitigate the fragility observed in many current implementations that overly rely on prompts and tools. Traditional designs, which abstract along implementation details, often lead to complexity and intertwined failures across components. In contrast, composing tasks enables clearer boundaries and facilitates the reuse of subtasks within larger workflows. As the field advances, this functional programming-inspired approach not only simplifies reasoning about complex AI systems but can also adapt to evolving technologies, ultimately enhancing the performance and resilience of AI solutions.