Show HN: AgentState – Open-source resilience and caching proxy for AI agents (github.com)

AgentState has unveiled an open-source resilience and caching proxy specifically designed for AI agents, aimed at preventing token wastage during execution failures. The tool intercepts calls from large language models (LLMs) and checkpoints execution states in an SQLite database, allowing users to pause, edit, and resume agent tasks from any point in the process. This functionality not only saves time but dramatically reduces costs associated with retries, as demonstrated in benchmark tests where AgentState achieved a 7,360x faster crash recovery time compared to standard agents and eliminated token costs on retries by achieving 100% cache hits. The significance of AgentState lies in its ability to enhance the reliability and efficiency of AI agents in production environments. With features like automatic checkpointing, a human-in-the-loop gateway for sensitive actions, and multi-model fallback capabilities, it mitigates risks associated with agent failures and ensures smoother operations. The open-source nature of the project encourages collaboration and integration within the AI ecosystem, allowing developers to adopt it easily into their workflows while benefiting from comprehensive features like session replay, real-time alerts for failures, and fine-tuning dataset export capabilities.