Pi Durable

Agentic applications now survive process crashes and can resume execution from their last checkpoint.

Earendil has released Pi 1.0, a stable version of its coding agent, alongside an experimental new package called Pi Durable. Pi Durable is designed as a "harness" to run long-running, durable, and malleable agentic applications. It supports agents that can operate anywhere with a JavaScript runtime, survive catastrophic failures, and manage infinitely long conversations, enabling multiple users to steer the same agents. Unlike the core Pi coding agent which focuses on single-user interaction on a local machine, Pi Durable is a framework for building any agentic application, including enhanced coding agents. It shares code and principles like minimalism and malleability with Pi, but operates independently to allow for exploration of new agent designs. The Pi Durable harness includes storage backends (memory, SQLite, JSONL), execution environments (like Node.js for file/shell access), and a system for managing tasks with checkpoints to ensure durability across crashes and restarts.

Pi Durable introduces a harness architecture that separates storage, model execution, and tool execution. This allows agentic applications to run across different environments, from local machines to remote VMs, and even across different machines for storage and execution. The design emphasizes malleability and extensibility, allowing developers to easily plug in new storage backends, execution environments, and tools, which can then be managed through a central registry. This modularity is key to building sophisticated, long-running agents that can adapt to complex workflows and user interactions.

The durability aspect is achieved through task checkpointing and request idempotency, ensuring that even if a process dies or a network connection is interrupted, the agent can resume its work without data loss or duplicate actions. This resilience is crucial for agents intended for continuous operation or complex, multi-step tasks. The ability to fork conversations and manage multiple concurrent agents within a single harness also signals a move towards more collaborative and multi-faceted AI agent systems, where different agents can interact and share history.