Pi 1.0 agent harness goes stable; Pi Durable ports it to TypeScript with crash-resumable state

Pi, the minimalist agent harness now part of Earendil, has reached version 1.0. A second release, Pi Durable, ports Pi to TypeScript and moves all stateful components out of the process so agents can resume after a crash. According to Latent Space's AINews roundup dated Oct 2, 2026, both releases reached the Hacker News front page the same day.

Pricing, licensing terms and benchmark results were not stated in the source material.

Pi 1.0: what changed

The source lists these features in the stable release:

• Codemode: native support for MCP and image models. The source also names "Jev" without explanation, and we could not verify what it refers to.
• Extension support for virtual models
• Deferred tool loading
• Cache warming for Anthropic models
• Mid-conversation system messages: transcript-aware changes to prompts and tools
• A new TUI theme, with full-screen mode on by default

Pi Durable: TypeScript port with externalized state

Pi Durable moves every stateful component of Pi outside the running process. Its documented capabilities:

• Crash survival: every step is recorded as a checkpointed task. If a process fails or restarts, agents and subagents resume from their last exact state.
• Portability: it runs on any JavaScript runtime, including Node, Bun and Cloudflare. Storage backends are pluggable (Memory, SQLite, JSONL), and execution can be remote or local.
• Concurrency: one harness can run multiple parallel, branching conversations, such as a main channel and separate threads, without blocking each other.
• Extensions: developers can bundle custom system prompts, tools, hooks and durable tasks into installable extensions. The example given is a multi-step checkout process with rollback.
• Context management: background compaction summarizes older messages to stay within token limits without pausing the agent.
• Multiplayer and state sync: application state, such as a to-do list, is stored as documents alongside conversation transcripts. Multiple users or UIs can connect, watch and steer the same agent.
• Hot-swapping: tool and extension code can be updated while the agent runs, and the next tool call picks up the new code.

These are descriptions of intended behavior from the project's materials as summarized by Latent Space. We have not seen independent testing of the crash-recovery or hot-swap claims.

Also in the same roundup

The issue's Twitter recap covered several model releases. Figures below are as reported there, and several are vendor-reported:

• GPT-6.1 Sol: Artificial Analysis reports $0.72 per Intelligence Index task at maximum effort, versus $1.04 for GPT-6 Sol and $3.26 for Astra. It attributes the gain to fewer turns and cheaper cache reads.
• Solar Mini 4 (Upstage): a text-only reasoning model, reported at 35B total and 3B active parameters, with a 1M-token context window and 262K maximum output. Weights are not released. Pricing is $0.10 input, $0.40 output and $0.01 cache-hit per 1M tokens. Artificial Analysis scores it 24 overall, 83% on long-context reasoning and 1% on Terminal-Bench 4.0.
• FLUX 3 Image (Black Forest Labs): native generation up to 4K with up to ten reference images. Commercial weights are available, and an open-weight variant is promised in coming weeks. BFL is running a 50% API discount through October 8 and has not given base prices in the cited posts.
• Gemini 4 Argon: announced by Google. Its coding quality is disputed between a Bloomberg report citing anonymous insiders and a rebuttal from a senior DeepMind engineer.

What this means

A 1.0 tag signals that Pi's interface is stable enough for others to build on. The more consequential release is Pi Durable. Durable execution, meaning checkpointed steps, resumable state and pluggable storage, is the main gap between agent demos and long-running production agents. Pi Durable brings that pattern to a lightweight harness rather than a heavyweight workflow engine.

Two features reflect current cost pressure. Anthropic cache warming and deferred tool loading both target token spend and latency, which matter more as agents run longer sessions with large tool sets.

The open questions are licensing, performance overhead from per-step checkpointing, and whether the TypeScript port reaches feature parity with the original. Until those are answered, treat the durability claims as design goals rather than verified guarantees.