Pi 1.0 ships native MCP and Code Mode, plus the Pi Durable agent harness
Pi has upgraded to version 1.0, with Code Mode and native MCP reaching stable release, and the team launched a new harness called Pi Durable the same day. According to co-founder Mario Zechner, Pi Durable uses an intent-execute-result "effect sandwich" and a versioned document store to persist user messages, model replies, tool results and task state so an agent can resume exactly where a crashed process left off. MCP, which Pi's founders had publicly criticized, is now built in with tool search and progressive disclosure; some users worry Pi is becoming bloated, while Armin Ronacher says the default context load has not increased.
Pi argues that any long-running agent needs durable execution; previously developers building assistants or Slack bots had to handle message storage, task scheduling and crash recovery themselves, the work Pi Durable aims to take over.
Pi 迎来了今年规模最大的一次更新：项目升级到 1.0，Code Mode 与原生 MCP 正式进入稳定版；同一天，团队还推出了一套全新的 Harness——Pi Durable。
这些变化已经酝酿了一段时间。过去一年，MCP 协议经历了六七轮调整，逐渐发展到真正可用的阶段；Pi Durable 的架构也前后经历了八轮迭代。一个曾被 Pi 公开拒绝的协议进入核心，一套反复打磨的长时 Agent 框架同时亮相，标志着 Pi 开始从终端编程 Agent 向完整的 Agent 基础设施延伸。
对开发者来说，Pi Durable 最直观的价值是省掉大量重复代码。过去在 Pi 上构建个人助理、Slack Bot 或其他 Agent 应用，开发者需要自己设计数据库保存消息，再处理 Agent 的启动、任务调度、进程崩溃和状态恢复。有开发者估计，Pi Durable 可以直接替代自己一半的代码，因为这些外围细节都由框架接管。
这也解释了为什么有人认为，Pi Durable 会成为许多 Agent 工作流和应用的底座。过去每家公司都在重复实现对话记录、任务持久化、故障恢复和会话分叉，Pi Durable 正好填补了这个空缺。它会保存用户消息、模型回复、工具执行结果和任务状态；进程崩溃后，Agent 可以从中断的位置继续。能够安全重放的操作会重新执行，可能产生重复副作用的操作则不会贸然重跑。
#Pi#Pi Durable#Code Mode#MCP#Earendil#Mario Zechner#Armin Ronacher#Jev#QuickJS