State-as-Files: A Manifesto for Multi-Session Agent Work (github.com)

A new manifesto, "State-as-Files," advocates for a structured approach to managing multi-session tasks with AI agents, emphasizing the inadequacies of traditional chat-based interactions. Currently, users often lose critical context when engaging with an AI after a session ends, resulting in inefficiencies and miscommunication on project goals. The proposal suggests treating project state like source code in Git, where everything is documented in files—ensuring that any agent, whether it's the same user or a colleague, can pick up where the last session left off without confusion. This framework specifies that each task should begin with a measurable goal documented in a file, which maintains a clear and immutable project history as sessions progress. Each agent interaction produces a structured "stage plan" that summarizes the prior work, enabling seamless continuation of projects that can span days or weeks. The approach also includes a verification process for artifacts produced during each phase, ensuring accuracy and accountability. By positioning state management as a core aspect of AI collaboration, "State-as-Files" aims to enhance trust, clarity, and efficiency in the increasingly complex landscape of long-term AI projects.