The lightweight Pi agent harness has reached version 1.0, adding native support for the Model Context Protocol, image models, virtual-model extensions and system-message changes during a conversation. The release is aimed at developers who want a small foundation for tool-using agents rather than a large application framework.
A related project, Pi Durable, ports the harness to TypeScript and moves stateful components outside the running process. Each step becomes a checkpointed task, allowing an agent or subagent to resume from the same state after a crash or restart. Storage options include memory, SQLite and JSON Lines files, while execution can be local or remote anywhere a JavaScript runtime is available.
Pi Durable can run branching conversations concurrently without blocking the primary session. Shared documents stored alongside transcripts let multiple interfaces or users observe and steer the same agent. Extensions can package prompts, tools, hooks and multi-step tasks, including workflows that need rollback. Tool and extension code can also be updated while an agent runs, taking effect on its next call.
These capabilities address durability and coordination in agent applications, but the release report does not provide independent reliability benchmarks. Teams still need to test checkpoint consistency, extension security and recovery behavior under their own failure conditions.