[Pi Reconstructs Agent Execution Layer: Long Tasks Moving Towards Persistent Operation] According to monitoring by Beating, Pi is reworking the Agent Harness. The latest Harness v3 specification has been completed and is undergoing final auditing. It has revamped task execution and storage, enabling Agents to resume from interruptions even if the process crashes midway or the Harness itself is upgraded. Pi is an open-source Agent Harness created by Mario Zechner, which initially gained attention due to its use by OpenClaw. The project was acquired by Earendil earlier this year, and Mario has joined the company to continue development. Currently, Pi has garnered over 60,000 GitHub stars, and MiniMax's latest MiniMax Code explicitly builds its Harness based on OpenCode and Pi. This time, v3 has completely overhauled the underlying architecture. Execution states are saved before model requests and tool calls, and results are written only after completion. Upon restarting, Pi can identify which tasks are completed, which can be rerun, and which have side effects that cannot be executed again. Conversation states, runtime states, and token costs are now stored separately, with added support for cross-version state migration. Both Claude Code and Codex are pushing Agents towards handling longer tasks. OpenAI has been specifically researching how to enable Codex to sustain long-duration work, and Anthropic has also stated that Claude Code is increasingly being used for long-running Agent tasks. Once Agents can operate continuously for several hours, the work done in the initial hours cannot be entirely lost due to a single crash. [Original Link]