Pi Durable

TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

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

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Earendil and the Pi community released Pi 1.0 and an experimental package called Pi Durable. The new framework is designed to support long-running agent conversations, persistent storage, different execution environments and recovery after process failures. Its broader reliability claims and planned capabilities have not been independently verified in the supplied report.

Earendil and the Pi community have released Pi 1.0 and an experimental package, Pi Durable, for developers building agent applications intended to run for long periods and recover after failures. Earendil says the framework is separate from, and does not replace, its existing Pi coding agent; instead, it provides a harness for managing conversations, tools, storage and execution environments.

Earendil describes a harness as the storage and machinery for running one or more conversations with language models, including the tools an agent can call and the environments in which those tools operate. In Pi Durable, a conversation is recorded as a transcript, while an agent consists of a model, its settings and its available tools. The framework treats model calls and tool executions as tasks.

The package includes memory, SQLite and JSONL storage options, along with a conformance suite and benchmarks intended for developers implementing their own storage backend. Earendil says the storage interface is small enough to adapt to systems such as a key-value store or Postgres. It also provides a Node execution environment for tools that need access to local files or a shell; developers can build other environments, including remote ones.

According to the report, a single process owns a storage instance, while other clients can attach to that process. With SQLite, the harness keeps active transcripts, live tasks and pending submissions in memory and leaves other data on disk until needed. Earendil says transcript compaction summarizes older messages before they exceed the model’s context window. These are descriptions of the project’s design, not independent test results establishing performance or reliability.

A Separate Layer for Persistent Agents

Pi Durable addresses a different operating model from Pi’s existing coding agent. Earendil describes that agent as running in a terminal on a user’s machine, controlled by one person; if it stops, the user can inspect what happened and tell it to continue. Pi Durable is aimed at applications that may need to remain available across restarts, run in different locations, or be steered by multiple people.

That distinction matters to developers deciding whether to build directly on the coding agent or use a framework designed to manage durable conversations and task execution. The storage and environment interfaces also offer potential flexibility: a team could keep the harness in one place while running tools elsewhere, or adapt storage to its own infrastructure. Those possibilities depend on implementation and testing; the announcement does not establish how the package performs under real-world failures or concurrent use.

Earendil calls Pi Durable experimental and invites the Pi community to help develop it. The release therefore represents an early project direction, not a claim that all the operational challenges of persistent agents have been solved.

How Pi Durable Relates to Pi 1.0

The announcement pairs Pi Durable with Pi 1.0, which Earendil says reflects extensive hardening, maintenance and active development of the existing coding agent. The company says Pi 1.0 remains focused on an agent used by one person in a terminal, and that this product is not being replaced by the new package.

Instead, Earendil presents Pi Durable as a framework for building a wider range of agentic applications, including coding agents. It shares code with the Pi coding agent, including the pi-ai project, and is intended to let developers explore new designs without disrupting the existing agent. Earendil says lessons from applications built with Pi Durable could later inform the coding agent if they prove useful.

The report describes the source code as approximately 15,000 lines, excluding tests. Earendil estimates that amount at about 150,000 tokens with GPT and 250,000 with Claude, while noting that builders generally would not need to load the entire codebase. These figures are the project’s own estimates and are not measures of capability or reliability.

“Pi Durable was built specifically for long-running, durable, and malleable agents that can run anywhere.”

— Earendil

Reliability Claims Await Testing

The supplied report does not state when the announcement was published, provide a package version number for Pi Durable, or list a license, installation instructions or a stable-release timeline. It also does not include independent tests showing how the framework behaves during crashes, redeployments, memory exhaustion or prolonged operation.

Earendil says the package is intended to survive failures and support long-running conversations, but the material provided ends during its discussion of crash recovery. It therefore does not specify the exact recovery guarantees, how in-progress tool calls are handled, or what data could be lost in a failure. The report also does not detail how multiple people steering one agent are coordinated, or provide benchmark results for the included storage options.

Support is described for JavaScript runtimes, with Earendil noting that the SQLite and JSONL storage code uses no Node APIs and can run with a small adapter in Bun or a Cloudflare Durable Object. The complete compatibility limits and production requirements are not given.

Community Testing and Development

Earendil is asking developers in the Pi community to help improve Pi Durable. The immediate next step described in the announcement is community exploration of the experimental framework, including its storage and execution-environment interfaces. Developers can use the stated conformance suite and benchmarks when building a custom storage backend, though the report does not give a schedule for further releases.

Earendil says insights from applications built on Pi Durable may feed back into the Pi coding agent when the team considers them valuable. Whether the experimental package advances to a stable release, and what reliability guarantees or additional deployment options it may offer, remain open.

Key Questions

What is Pi Durable?

Pi Durable is an experimental framework from Earendil for building agentic applications. It provides storage and machinery to manage model conversations, tools and the environments where those tools run.

Does Pi Durable replace the Pi coding agent?

No. Earendil says the package does not replace the Pi coding agent. The existing agent remains focused on a person using it in a terminal, while Pi Durable is a separate framework for building agent applications.

What storage options does the package include?

The announcement lists memory, SQLite and JSONL storage. It also describes a conformance suite and benchmarks for developers who want to build their own storage backend.

Can Pi Durable recover an agent after a crash?

Crash recovery is one of the package’s stated goals. Earendil says it is designed for agents to resume after a process dies, but the supplied report does not spell out recovery guarantees or provide independent test results.

Is Pi Durable ready for production use?

Earendil labels Pi Durable experimental. The announcement does not provide a stable-release date or enough testing information to establish production readiness.

Source: hn