According to Beating monitoring, Pi is reworking the Agent Harness. The latest Harness v3 specification has been completed and is undergoing final audit. It reworks task execution and storage, allowing an Agent to continue from where it left off even if the process crashes mid-run or the Harness itself upgrades. Pi is an open-source Agent Harness created by Mario Zechner, which first gained attention because OpenClaw used it. The project was acquired by Earendil this year, and Mario joined the company to continue development. Pi now has over 60,000 GitHub stars, and MiniMax's latest MiniMax Code explicitly builds its Harness on OpenCode and Pi. V3 directly reworks the underlying architecture. Model requests and tool calls now save execution state before they are made, and write results after completion. After a restart, Pi knows which tasks are done, which can be rerun, and which have side effects and cannot be executed again. Conversation, run state, and token costs are also saved separately, with cross-version state migration added. Claude Code and Codex are both pushing Agents toward longer tasks. OpenAI has already studied how to make Codex continuously complete long-running tasks, and Anthropic also states that Claude Code is increasingly being used for long-running Agent tasks. Once an Agent can work continuously for several hours, the first few hours should not be wasted just because of a single crash.