Earendil has released Pi 1.0, the first stable version of its agent harness that hundreds of thousands of developers already use each week to run coding agents from major AI providers. Alongside it, the company introduced Pi Durable, an experimental new package aimed at letting people build long-running AI agent applications that work outside the terminal. Both are notable because they represent a deliberate bet against the fast-churning “move fast and bolt on features” culture common in agentic AI tooling.

What Pi actually is

Pi is described as an agent harness: software that wraps around large language models to let them act as autonomous or semi-autonomous coding agents, calling tools, running commands, and completing multi-step tasks. Rather than being tied to one AI provider, Pi has supported the latest models from every major provider, making it a flexible, provider-agnostic base. Many developers have adopted it as a daily driver for coding work, and it has also served as what Earendil calls a “supermalleable substrate” for building other agentic applications on top of it.

Why minimalism is the core design principle

The defining trait of Pi, according to Earendil, is restraint. The team says it deliberately waits until a new agentic technique or pattern has proven itself before adopting it, weighing real functionality against the complexity it would add to the codebase. This is a direct contrast to much of the agentic tooling landscape, where new frameworks, protocols, and integration patterns appear constantly but often do not last. Earendil points to its own recent discussion of Codemode and MCP (Model Context Protocol) as an example of how it evaluates whether a trend is durable enough to bring into Pi, rather than chasing every new idea.

This matters for developers because tool sprawl and unstable dependencies are a real cost in agentic development. A harness that changes its interface or behavior every few weeks is hard to build a business or workflow around. Pi 1.0 is explicitly positioned as the result of months of filtering ideas: the post notes that many proposed features “were thrown up against the wall,” and far more were rejected than were kept.

What is new in version 1.0

The 1.0 release folds in a set of features that Earendil says it considered carefully over many months before committing to them. The source material does not enumerate the specific feature list in detail, but frames the release as a “major step forward” that still preserves Pi’s core simplicity. The emphasis throughout is that scale and new capability were added without compromising the minimal footprint that made Pi usable and dependable for its existing base of users submitting issues and pull requests over time. That ongoing community feedback loop is cited as a key part of how Pi was hardened into something “people and businesses can depend on.”

Pi Durable: taking the agent out of the terminal

Perhaps the more significant structural change is Pi Durable, a new, separate experimental package rather than a feature bolted onto core Pi. Earendil says it recognized that many people wanted to use Pi in ways that did not fit its original shape: reachable from different surfaces beyond the terminal, and capable of supporting longer-running conversations and tasks rather than single coding sessions.

Instead of stretching core Pi to cover these use cases and risking its minimalism, Earendil built Pi Durable as a distinct substrate. Pi Durable is meant for building long-running agentic applications where both the builder and the end user can steer the underlying AI model with fine-grained control, even across extended, multi-step interactions. The company says it shares Pi’s founding principles, minimalism and “supermalleability” (the ability to be reshaped for many different applications) but extends those ideas into new territory, specifically durability over time and reachability across surfaces other than a terminal window. A companion technical writeup from team member Mario goes into more detail on the design choices behind Pi Durable, though the summary here focuses on its purpose rather than implementation specifics.

How to try it and what to expect

Pi 1.0 is available now, and Pi Durable ships today as an experimental package, meaning it should be expected to change and mature rather than treated as a finished product. Both packages are released under the MIT license, making them free to use, modify, and redistribute, including in commercial projects. Documentation lives at pi.dev, and the source code is hosted at github.com/earendil-works/pi for developers who want to inspect the implementation, file issues, or contribute pull requests, which Earendil notes has already shaped much of Pi’s development to date.

For developers already using Pi as a coding agent, the practical takeaway is that the 1.0 label signals a level of stability intended for production and business use, not just experimentation. For those interested in building agent-powered products that run longer than a single terminal session, such as chat assistants, background task runners, or multi-step workflow tools, Pi Durable is the new entry point worth evaluating, with the caveat that its experimental status means APIs and behavior may still shift.

Frequently asked questions

Is Pi tied to a specific AI model provider?
No. The source states Pi already runs the latest models from every major provider, positioning it as a general-purpose harness rather than a single-vendor tool.

What is the difference between Pi and Pi Durable?
Pi is the original minimal, terminal-centric coding agent harness now at version 1.0. Pi Durable is a new, separate experimental package for building long-running agentic applications that can be reached from surfaces other than the terminal and that support extended conversations and tasks.

Is Pi open source?
Yes, both Pi and Pi Durable are MIT licensed, with code available on GitHub and documentation at pi.dev.

Source: Pi 1.0