Agent Framework

Pi 1.0

2026-10-02 03:35:01

Pi 1.0 launches as Earendil introduces Pi Durable for long-running agents

Earendil has officially released Pi 1.0 and, alongside it, introduced an experimental project called Pi Durable. The Pi 1.0 release largely formalizes capabilities that had already gone live in recent updates, while the main new product line is Pi Durable, a lower-level framework built for long-running agents. According to the release note summarized by BlockBeats, Pi Durable continuously writes task state to storage, allowing jobs to resume after a process crash, container restart, or even a computer sleep event. The framework breaks model calls and tool calls into recoverable tasks and saves checkpoints at every step. If a model request is interrupted, it can be retried. Tool execution is only rerun automatically when it is confirmed to be safe to do so. For tools such as deployment or payments, where repeating an action could create duplicate operations, the interrupted state is handed back to the model for handling. Pi Durable also supports multiple simultaneous conversations, queued messages, saved sub-agent states, and automatic compression of overly long context. Earendil said Pi Coding Agent will remain a single-user terminal tool, while more complex persistent agent use cases will be handled by Pi Durable. Both projects use the MIT license.

20

open-source

2026-09-01 13:53:40

Open-Source Agent Framework Pi Reaches 100K GitHub Stars, v2 on the Horizon

Pi, an open-source agent harness, has surpassed 100,000 stars on GitHub. The team also teased the upcoming Pi v2. Over the past month, the team has been rewriting the underlying agent runtime, evolving from Harness v2 to v3, adding task persistence, crash recovery, and parallel execution lanes. The new Harness aims to allow agents to resume from a safe state after a crash and support concurrent tasks on the same session history.

780

DeepSeek

2026-08-14 02:52:25

DeepSeek open-sources Harness agent framework as GitHub stars near 60,000 within hours

DeepSeek open-sourced its agent framework, DeepSeek Harness, on the night of Aug. 13, releasing the v0.1 developer preview under the MIT license. According to the source material, the project picked up nearly 60,000 GitHub stars within just a few hours, drawing strong community attention. The framework is positioned against Anthropic’s Claude Code and OpenAI’s Codex, but it takes a different route. DeepSeek describes Harness with the formula "Model + Harness = Agent," where the model handles reasoning and Harness manages tool calls, task state, and the execution loop. The company’s main design pitch is "Everything is a plugin," with models, tools, conversation state, filesystem, sandbox, agent loop, orchestration logic, and even the web interface built on a plugin microkernel called Cordis. Harness ships with four preset modes, supports connections to DeepSeek, Anthropic, OpenAI, and about 40 model providers in total, according to the official description cited in the source. The article says it may be worth trying for teams building agent products or infrastructure, especially because it preserves each step of a conversation and exposes which tools and models an agent is using. At the same time, the source flags unstable APIs, incomplete documentation, a web UI-focused approach, and supply-chain risk during the preview stage.

1800

Qwen3.6-27B

2026-07-22 13:55:13

Alibaba Open Sources Qwen3.6-27B: 27B Model Matches Claude 4.5 Opus in Agent Tasks, Cost Cut by 14x

Alibaba's Qwen team open-sourced Qwen3.6-27B on April 22, a 27B dense model that ties Claude 4.5 Opus on Terminal-Bench 2.0 and beats its 397B MoE predecessor on SWE-bench. Quantized to 16.8GB, it runs on consumer hardware, becoming a local brain for OpenClaw and Hermes Agent.

490