Pi 1.0 Release: A Minimalist Agent Harness for AI Orchestration

Earendil has officially released Pi 1.0, a minimal and extensible agent harness designed to provide a stable substrate for building and wielding AI agents. The release marks a transition from an experimental phase to a hardened piece of software intended for both individual developers and businesses to depend on.

Core Features of Pi 1.0

Pi 1.0 introduces several key technical enhancements aimed at increasing the flexibility and efficiency of agentic workflows:

Codemode: Native support for the Model Context Protocol (MCP), as well as support for non-LLM models (such as Jev) and image models.
Virtual Model Extensions: Support for extensions that allow the creation of virtual models. This enables complex routing logic, such as using one model for planning (e.g., Claude Opus) and another for implementation (e.g., GPT-6 Luna), with a router model (like Jev) deciding when to switch between them.
Deferred Tool Loading: Optimizes startup and resource usage by loading tools only when required.
Anthropic Cache Warming: Integrated support for cache warming to reduce latency and costs when using Anthropic models.
Transcript-Aware Prompting: Support for mid-conversation system messages, allowing for prompt and tool changes that are aware of the current conversation transcript.
TUI Improvements: A new TUI theme and full-screen mode enabled by default to enhance the terminal user interface experience.

Pi Durable: Long-Running Agentic Applications

Recognizing that a terminal-based coding agent does not fit every use case, Earendil has introduced Pi Durable, an experimental package. Pi Durable is designed as a substrate for building long-running agentic applications that require durability and the ability to be steered with high dexterity. It extends Pi's core principles of minimalism and malleability into dimensions that allow agents to operate outside the terminal and across different surfaces.

Technical Implementation and Ecosystem

Pi is MIT licensed and available via a simple installation script for macOS/Linux and PowerShell for Windows. For developers building on the SDK, the Pi Durable ecosystem includes packages such as @earendil-works/pi-durable, @earendil-works/pi-ai, and @earendil-works/chord.

Community Insights and User Feedback

Community discussion on Hacker News reveals a strong preference for Pi's minimal system prompts, which users report makes it more compatible with local models and limited context windows compared to other harnesses.

Performance and Extensibility

Users have praised the harness for being "hackable" and vendor-agnostic. One developer noted that the Pi SDK allows for the creation of custom harnesses for platforms like Slack, though they mentioned the complexity of managing JSONL session files in Kubernetes environments.

"I initially used Codex, but Pi seems more hackable, and I like that it's vendor-agnostic by default."

Critical Perspectives

Some users expressed skepticism regarding the "minimalism" branding, citing the project's dependency on Node.js/NPM and its significant lines of code (SLOC). Others criticized the choice of JavaScript/TypeScript for a CLI tool with high interactivity requirements, suggesting that languages like Rust would be more performance-efficient.

Security and Sandboxing

Users have shared various strategies for securing the agent's access to their systems, with some utilizing tools like bubblewrap (via a wrapper called sbh) to ensure the agent only has access to specific directories and network capabilities.

Comparison to Other Tools

While some users find Pi superior to alternatives like Claude Code or OpenCode due to its extensibility, others suggest that the simplicity of a single-agent setup in Pi (without extensions) is often faster and more token-efficient than the multi-agent orchestrations found in other commercial harnesses.