Earendil has released Pi 1.0 alongside an experimental package called Pi Durable, a framework for building long-running, fault-tolerant AI agent applications.
Pi 1.0 focuses on the original Pi coding agent—a single-user tool that runs in a terminal on a remote machine. Pi Durable, by contrast, is designed to bring agent technology to broader use cases. According to Earendil, Pi Durable is built for applications that need to run anywhere, survive crashes, support multiple concurrent conversations, and allow multiple users to steer the same agents.
What is Pi Durable?
Pi Durable functions as a “harness”—a system combining storage and machinery to run one or more conversations with large language models in parallel. A conversation is recorded as a transcript between a user and an agent (the LLM plus its settings). The harness provides tools the models can call and execution environments where those tools run.
According to Earendil, the source code is approximately 15,000 lines—about 150,000 tokens in GPT terms or 250,000 with Claude—making it small enough for agents to understand and build upon.
Key Features
Pi Durable ships with multiple storage backends including memory, SQLite, and JSONL formats. The SQLite implementation keeps only the active working set in memory (active transcripts, live tasks, and pending submissions), with everything else remaining on disk. Even conversations with tens of thousands of messages fit within memory constraints because the system uses compaction to summarize older messages before they overflow the model’s context window.
The framework can run anywhere there is a JavaScript runtime, including on Bun or inside Cloudflare Durable Objects. Tools requiring files or shell access receive them from pluggable execution environments. Pi Durable ships a Node execution environment for local file access, but the interface is designed so developers can implement remote execution environments, allowing the harness to run on one machine while tools execute on another.
Crash Resilience
Every step of a run is stored as a task with a checkpoint. If the process dies—whether from a laptop sleep, container redeployment, or memory issues—a new process can open the same storage, find unfinished tasks, and continue from the last checkpoint. Partial model responses are marked as aborted; interrupted tool calls either rerun if safe or inform the model of the interruption.
Concurrent Conversations
One harness can run multiple conversations simultaneously without blocking. Conversations can start fresh or fork from another conversation at any point in its transcript, viewing parent history without copying it.
Earendil stated that Pi Durable shares principles of minimalism and malleability with the Pi coding agent, and does not replace it. The company plans to use lessons learned from building applications on Pi Durable to inform future improvements to the original Pi coding agent.
Key facts
- Earendil released Pi 1.0 and the experimental Pi Durable framework for AI agents
- Pi Durable is designed for agents that run anywhere, survive crashes, support multiple users, and handle parallel conversations
- The framework uses pluggable storage backends (memory, SQLite, JSONL) and execution environments
- Pi Durable keeps only active working sets in memory, storing everything else on disk
- The harness can continue from checkpoints after crashes, marking partial responses as aborted