Earendil ships Pi 1.0 and a TypeScript port called Pi Durable, both MIT licensed
Earendil said on 1 October that Pi, its minimal agent harness, is now at version 1.0, and shipped a second experimental package called Pi Durable that ports the same ideas to TypeScript for long-running agents.

Image: Earendil
Why it mattersA team using Pi now has a stable base to build on, and a separate TypeScript runtime that lets the same agent survive a process crash and run on a Cloudflare worker instead of only in a terminal.
A coding agent that only runs in a terminal is one process away from losing the whole session. Earendil said on 1 October that it is shipping two things that answer that in different ways: Pi 1.0, a stable release of its minimal terminal agent, and Pi Durable, a separate TypeScript package that lets the same agent survive a crash and run on a server.
Both are MIT licensed. Pi 1.0 is the stable release of the Pi coding agent after what Earendil describes as "countless hours of hardening". The company said in its post that "hundreds of thousands of people around the world use Pi every week". That is Earendil's own number and the post does not break it down further.
What is new in Pi 1.0
Earendil lists seven changes in the 1.0 release. The headline one is Codemode, which gives Pi native support for the Model Context Protocol and for non-LLM models like Jev and image models. Codemode is the JavaScript sandbox Earendil added to Pi on 29 September to let the agent compose tool calls on the trusted side of the harness, before the model sees the result.
The rest are smaller: extension support for virtual models, deferred tool loading, cache warming for Anthropic models, mid-conversation system messages that let a prompt or tool list change inside a running conversation, a new terminal theme, and full-screen mode on by default.
The example the post gives for virtual models is a single Pi session that routes planning to Claude Opus and implementation to GPT 6 Luna, with Jev deciding when to hand over between them. Pi then prints the cost broken down by model and by cache use.
What Pi Durable actually is
Pi Durable is a separate experimental package that ports Pi's harness to TypeScript. The whole source, without tests, is about 15,000 lines, which Earendil says comes out to about 150,000 GPT tokens or 250,000 Claude tokens, so an agent can read the whole thing when building on it.
The package ships three storage backends out of the box: in-memory, SQLite, and JSONL on disk, plus a conformance suite and benchmarks. The SQLite and JSONL backends use no Node APIs, so they also run on Bun and inside a Cloudflare Durable Object. Every step the harness runs is recorded as a checkpointed task, so a process that dies or restarts picks up its agents and subagents from their last exact state. Tool and extension code can be swapped while the agent is running, and the next tool call picks up the new code.
Several things in Pi Durable are not in Pi 1.0 at all: multiple parallel branching conversations inside one harness, multiplayer state sync so several people can watch and steer the same agent through documents stored next to the transcript, and automatic background compaction that summarises old messages before they overflow the context window.
Earendil says Pi Durable does not replace the Pi coding agent and that lessons from building on it "will flow back into Pi the coding agent as they prove themselves valuable". The split keeps the terminal agent small for the person using it alone and gives a separate runtime for anything that needs to outlive the terminal: a chat UI, a long-running job, an agent reached from a browser. A team already on Pi has a stable 1.0 to pin to and gets the durable runtime from the same codebase, sharing the pi-ai package.
Source
Pi 1.0 and Pi Durable, Earendil, 1 October 2026. Discussion on Hacker News.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

