Earendil Launches Pi 1.0 Agent Harness With Durable Session Support
October 5, 2026 · 3 min read
Earendil announced on October 1 2026 that it released version 1.0 of its Pi agent harness. The update sets terminal UI fullscreen display as default and improves Codemode tool invocation via JavaScript. The release also introduces Pi Durable, an experimental foundation that saves agent execution state and enables process resumption after interruption.
Pi functions as an agent harness that manages interactions with AI models and controls tool execution. It can operate as a terminal coding agent or be embedded in applications through an SDK. Core functionality remains minimal while extensions, skills, prompt templates and themes allow customization for user workflows. Subagents and plan modes can be added via optional packages.
Earendil states that months of iterative improvements based on user feedback have produced stable software trusted by individuals and enterprises.
The terminal UI now defaults to fullscreen display and can switch to a scroll history view that mirrors native terminal behavior. Users may also configure startup to show only a header containing version information and key operation guidance.
Codemode reduces prompt token consumption by approximately 40 percent when requesting GPT 5.6 with default tools enabled. The change shortens tool descriptions and makes detailed API documentation available on demand. A new models generateImages function lets Codemode scripts create images using session authentication information and display results. Image generation costs are included in overall session expenses.
Error handling receives enhancements. Invalid tool names trigger suggestions of similar names and incorrect argument formats for image generation APIs now display correct syntax. OAuth authentication for MCP servers now stores credentials per server name and URL combination, allowing multiple accounts to share a URL.
Pi Durable provides a framework for persistent agent development. It targets long running tasks such as coding or travel planning and supports operation across multiple screens and clients. The framework shares the pi ai library with Pi for model access. Armin Ronacher, a Flask developer involved in Earendil development, positions Pi Durable as a foundation for future agent projects.
Pi includes session persistence but expects user confirmation to resume after interruption. Pi Durable manages tasks as checkpoints storing conversation history and state in SQLite or JSONL files. Upon restart the system reopens the data store and resumes from the last checkpoint. Replayed model requests occur only for safely restartable tool actions.
The storage mechanism and execution environment remain customizable. Tools can run on separate machines distinct from the Pi Durable process.
For lengthy conversations the system summarizes older messages in the background while preserving original content for subsequent turns. If the next request exceeds context limits the summary completion blocks progress.
Multiple conversations can run concurrently and branch from any historical point. New participants can join existing sessions and interact with current state including conversation history generated responses and active tool outputs.
Extension code can replace active tool implementations without terminating the conversation. New tool calls adopt the updated code starting with the next invocation.
Both Pi 1.0 and Pi Durable are released under the MIT license. Pi is available through the official website and Pi Durable includes sample code in its public repository for experimental development.
Earendil explains the adoption of TypeScript in Pi Durable stems from ease of project initialization while acknowledging potential future migration to Rust or assembly languages.