TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Earendil and the Pi community released Pi 1.0 and an experimental package called Pi Durable. The framework is designed for developers building long-running agent applications that can use persistent storage and different execution environments; it does not replace the Pi coding agent.
Earendil and the Pi community have released Pi 1.0 and, alongside it, an experimental package called Pi Durable for building long-running AI agent applications. Earendil says the framework is intended to support persistent conversations, recovery after failures and agents that can be accessed through different interfaces, while leaving the existing Pi coding agent unchanged.
Earendil describes Pi Durable as a framework for building agentic applications, including coding agents, rather than a replacement for Pi’s existing coding agent. The company says that agent is designed to run on a user’s local or remote machine in a terminal, generally under one person’s direction. Pi Durable is meant to support broader arrangements, including multiple people steering an agent and agents that continue running over long periods.
The framework defines a harness as storage plus the machinery for running conversations with language models, including the tools agents call and the environments where those tools execute. Its storage options include memory, SQLite and JSONL. Earendil says developers can implement the storage interface for other systems, and that the SQLite and JSONL implementations avoid Node APIs so they can be adapted for runtimes such as Bun or Cloudflare Durable Objects.
Pi Durable also includes a Node execution environment that gives tools access to local files. Developers can implement other environments, including remote ones, allowing the harness and the tools’ execution environment to run on separate machines. Earendil says the SQLite-backed harness keeps active transcripts, live tasks and pending submissions in memory while leaving other data on disk; it also says transcript compaction summarizes older messages as conversations approach a model’s context limit.
A Framework for Persistent Agent Runs
Pi Durable addresses a practical limitation of agent prototypes: a process or host can stop while an agent still has work to do. By combining persisted conversation data with task execution and restart-oriented design, the package aims to let developers build applications that can resume work rather than relying on a single uninterrupted terminal session. That matters for services expected to operate over longer periods or across changing infrastructure.
The architecture also separates the harness from the environment where tools run. Developers could, for example, keep conversation and task management on one machine while running file or shell tools elsewhere. Earendil presents this flexibility as a way to make agents usable across different deployment settings and interfaces. These are design goals described by the project; the announcement does not provide independent reliability results or production performance measurements.
For Pi users, the distinction is relevant: the company says the coding agent remains focused on the individual, terminal-based workflow, while Pi Durable gives developers a separate place to test broader agent designs. Earendil says useful lessons from those applications may later inform the coding agent, but does not promise that particular features will be transferred.
As an affiliate, we earn on qualifying purchases.
How Pi Durable Relates to Pi 1.0
Pi 1.0 is presented by Earendil as a milestone for the existing coding agent, following what the company describes as extensive hardening, maintenance and active development. The report characterizes Pi as a foundation for further building, while acknowledging that development will continue. It supplies no independent assessment of the 1.0 release.
Pi Durable shares code with the coding agent, including the pi-ai package, and is intended to share its principles of minimalism and malleability. Earendil says keeping it as a separate experimental package lets the team explore agent-application designs without disrupting the established coding-agent product. The report estimates the source code at about 15,000 lines excluding tests, while noting that developers generally would not need to read all of it to build on the framework.
The release description includes memory, SQLite and JSONL storage options, a storage conformance suite and benchmarks intended for developers testing their own backends. It also describes conversations as transcripts and model calls and tool executions as tasks. Those details outline the package’s intended architecture, but the announcement does not document a full independent evaluation of its behavior under failure.
“Pi Durable was built specifically for long-running, durable, and malleable agents that can run anywhere.”
— Earendil, in its Pi Durable announcement
As an affiliate, we earn on qualifying purchases.
Reliability Claims Await Testing
Pi Durable is described as experimental, and Earendil’s release report does not provide independent testing, uptime figures or benchmarks showing how reliably agents recover from crashes in real deployments. It describes storage choices and architecture, but does not establish that every failure scenario will preserve work or resume without intervention.
The announcement also does not specify service-level guarantees, security controls, the exact limits of simultaneous access, or which interfaces for multiple human operators are currently available. Although the source says the framework is intended to survive failures and support long-running conversations, the supplied report ends partway through its discussion of crash recovery. Further implementation details are therefore not available in the material provided.
No calendar date, pricing, licensing details or adoption figures are included in the source report. Those points should not be inferred from the release announcement.
SQLite database for AI applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Further Development and Community Testing
Earendil is asking developers in the Pi community to try Pi Durable and help improve it. The project’s stated next step is continued experimentation with the framework, including the storage and execution-environment interfaces, as developers build agentic applications with it.
Readers evaluating the package can look for additional release documentation, implementation examples and test results, particularly evidence about recovery after process or infrastructure failures. Earendil says lessons from Pi Durable may inform the Pi coding agent if they prove valuable; the announcement does not give a schedule or identify specific changes planned for that product.
JSONL storage for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Pi Durable?
Pi Durable is an experimental framework from Earendil for building agentic applications, including coding agents. It provides storage and machinery for running model conversations, tools and their execution environments.
Does Pi Durable replace the Pi coding agent?
No. Earendil says Pi Durable is a separate framework and does not replace the Pi coding agent, which remains focused on a terminal-based workflow.
Which storage options does it include?
The announcement lists memory, SQLite and JSONL storage. Earendil also says developers can implement the storage interface for other backends.
Has Pi Durable been proven to recover from every failure?
No such proof is provided in the release report. Earendil describes recovery as a design goal, but the package is experimental and the announcement includes no independent reliability results or production guarantees.
Where can Pi Durable run?
Earendil says it is designed for JavaScript runtimes. The announcement describes Node execution tools and says the SQLite and JSONL storage code can be adapted for Bun or Cloudflare Durable Objects, among other possible environments.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
