Type: url
Original: https://jimmysong.io/zh/book/ai-handbook/runtime/harness/
Status: generated
Language: zh
Summary
This chapter of Jimmy Song’s online book on building AI agents defines the harness as the runtime substrate between the model and the application that mediates observation, action, feedback and completion, and lists six core responsibilities (observation, action, feedback, completion/verification, permission and memory) whose completeness separates a harness from a framework. It contrasts the framework and harness paradigms, presents Hermes Agent (always-on), Claude Code (developer-facing) and OpenAI Codex (cloud-asynchronous) as three productized forms, and positions the harness as the single-agent execution layer beneath a multi-agent Agent Runtime. It closes with the security risks of always-on agents, such as prompt injection, confused-deputy attacks and permission escape, and argues for least-privilege design.