OpenHarness: Claude Code-style agent architecture

This Japanese-language overview presents OpenHarness as an open source Python framework from the University of Hong Kong's data science lab for building a coding assistant with behavior similar to Claude Code. The explanation focuses on the infrastructure around a model: tools let it act, memory retains project context, and permissions control operations.

The speakers describe an AI agent loop that reads code, makes changes, runs tests and uses the results to decide what to do next. They also discuss exponential backoff for API errors and usage tracking. Project instructions in CLAUDE.md and persistent notes in MEMORY.md provide context across sessions, with automatic summarization when that context grows too long.

Permission settings allow, request confirmation for or deny operations. Hooks can check actions before tool execution or record results afterward. For larger tasks, the framework reportedly creates subagents and tracks their roles and progress in a team registry.

The provider discussion covers Anthropic-compatible and OpenAI-compatible interfaces. The stated stack requires Python 3.10+ and uses React with Ink for the terminal interface. The speakers describe a one-line curl installer and an MIT license. This is an architecture overview; it does not demonstrate how to run models locally or establish offline operation.