Pi Refactor Agent Execution Layer: Long Task Transition to Persistent Execution

According to Motion Insight Beating Monitor, Pi is undergoing a redesign of the Agent Harness. The latest Harness v3 specification has been finalized and is undergoing final audit. It has revamped task execution and storage, allowing the Agent to resume from the interruption point even if it crashes midway through the process or if the Harness itself is upgraded.

Pi is an open-source Agent Harness created by Mario Zechner, initially gaining popularity due to its use in OpenClaw. The project was acquired by Earendil this year, with Mario also joining the company to continue development. Pi now has over 60,000 GitHub Stars, and the latest MiniMax Code from MiniMax clearly builds on OpenCode and Pi to construct the Harness.

This time, v3 has directly redesigned the underlying architecture. The execution state is now saved before model requests and tool invocations, and the results are written upon completion. Upon restart, Pi can identify tasks that are completed, ones that can be rerun, and those with side effects that cannot be repeated. Conversations, execution states, and Token costs are now stored separately, with the addition of cross-version state migration.

Claude Code and Codex are both pushing the Agent toward longer tasks. OpenAI has been specifically studying how to keep Codex working continuously on long-term tasks, and Anthropic has noted that Claude Code is increasingly being used for long-running Agent tasks. An Agent that can work continuously for hours shouldn't lose all progress due to a single crash in the initial hours.