Pi 1.0 arrives with leaner tools and experimental Pi Durable
Earendil releases Pi 1.0 and an experimental framework for persistent agents, with leaner tool prompts and explicit limits on crash recovery.
Earendil released Pi 1.0 on October 1 alongside Pi Durable, an experimental framework for AI agents that keep working across longer conversations and interruptions. The paired release gives developers a stable coding tool and a separate foundation for building persistent applications.
The company announcement describes Pi as a deliberately compact agent harness. That is the software connecting a model to tools and conversation history. Both Pi and Pi Durable are available under the MIT license.
What changes in Pi 1.0
The version 1.0 release notes make the newest changes clearer. The terminal interface now opens in fullscreen mode, with an option to retain ordinary terminal scrolling. Codemode scripts can generate images using the session’s configured credentials, and tool errors include more useful recovery guidance.
Pi also reduces the prompt overhead of Codemode. Its example for a GPT 5.6 request with default tools falls from roughly 5,300 tokens to 3,300. That is a reduction in the instructions sent to the model, rather than a measured saving across every complete task.
Some headline capabilities arrived before the 1.0 milestone. The dated changelog records MCP support, Codemode and virtual model routing on September 29. Cache warming and changes to prompts during a conversation appeared on September 19. Version 1.0 brings that work together with further improvements.
Pi Durable tackles recovery
The Pi Durable engineering announcement describes checkpointed tasks that can resume when a new process reopens the same persistent storage. SQLite and JSONL support that design. A memory backend is also available, and tool execution can run on a different machine.
| Product | Main role | Release status |
|---|---|---|
| Pi 1.0 | Coding agent operated from a terminal | Stable release |
| Pi Durable | Framework for persistent agent applications | Experimental package |
Recovery has an important limit. An interrupted tool call is repeated only when its tool declares that replay is safe. Otherwise the interruption is reported to the model. A request identifier prevents a retried submission from being submitted twice, but developers still need to design external actions around their own failure conditions.
The framework can run several conversations concurrently. It does not include built in subagents, although the engineering guide demonstrates how to create them using separate conversations and tools.
What developers should evaluate
For a team adopting persistent agents, the practical test is recovery under realistic failures. Try stopping a process during a read, a model response and an external write. Check what resumes, what needs review and whether the resulting history remains understandable.
ByteForward has not independently tested these release claims. Pi Durable remains experimental and its API may still change. Our Claude Code Mods coverage examines another approach to extending an AI coding workflow.
Original AI generated conceptual illustration of an agent workflow and saved task state. Created for ByteForward.
