Lessons from Replit and Tiger Data on Storage for Agentic Experimentation

Replit and Tiger Data have independently developed storage architectures to support agentic experimentation, each arriving at the conclusion that forkable state is essential for such environments. Replit's system, built on Google Cloud Storage, utilizes a three-tier architecture that accommodates agent sandboxes by leveraging immutable 16 MB objects to manage storage, suitable for localized access patterns but challenging for workloads requiring fine-grained, random access. In contrast, Tiger Data's Fluid Storage employs a block-based approach that manages individual 4 KB blocks, facilitating both agentic experimentation and operational workloads with lower latency and reduced read/write amplification. While Replit's architecture is designed for internal use within its platform, Tiger Data's solution serves as a shared infrastructure available to all customers. Both systems highlight a shift in infrastructure requirements, driven by the need for continuous and stateful experimentation in agent-based development environments.