这是一篇由原始材料转换而来的阅读页，保留了源文件的主要结构，并补充了可追溯的来源说明与链接。

Sources

Primary

1. George (@odysseus0z), Harness Engineering Is Cybernetics
   X article: https://x.com/odysseus0z/article/2030416758138634583
   Related post: https://x.com/odysseus0z/status/2030416758138634583

2. Anthropic Engineering, Effective harnesses for long-running agents
   https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

3. OpenAI, Harness engineering: leveraging Codex in an agent-first world
   English: https://openai.com/index/harness-engineering/
   Chinese: https://openai.com/zh-Hans-CN/index/harness-engineering/

Key framing extracted from George

• Harness engineering is analogous to cybernetics / control systems.
• Humans shift from direct operation to steering via specs, feedback loops, and constraints.
• LLMs close a feedback loop at a higher layer of abstraction than compilers/tests/linters alone.
• The hard problem is calibration: making judgment machine-readable.

Key framing extracted from Anthropic

• Long-running agents fail when they one-shot, lose continuity, or declare victory too early.
• Minimal durable artifacts help: init.sh, progress log, feature list, git history.
• Each session should make one verified incremental change and leave a clean handoff.

Key framing extracted from OpenAI

• The repo itself should become the agent’s operating environment.
• AGENTS.md should be a map, not an encyclopedia.
• Docs, plans, observability, and architecture rules should be versioned in-repo.
• Constraints should be mechanical; governance should be continuous.