Tag
Harness Engineering
Harness engineering is the layer around an LLM that shapes how it runs: orchestration, state isolation, tool access, memory handling, and recovery paths. It matters because these controls decide whether agents can handle long tasks reliably and ship as production systems.
4 articles
From Prompting to Harness Engineering
OpenAI says one team shipped a 1M-line product with 3 engineers and Codex, merging about 1,500 PRs in 5 months.
Harness Engineering for Long-Running Multi-Agent Systems
A context-reset design keeps each Claude Code run clean, turning Planner output into JSON so Generator stays focused on the task.
Claude Code 里的 Harness Engineering 思路
Claude Code 把 MCP、Skills、Hooks 和 Subagents 直接暴露给用户，展示了 Anthropic 如何把 Harness Engineering 做进产品。
Harness Engineering: From Bridle to Operating System, The Missing Link in AI Agent Reliability
Harness Engineering is the discipline of designing external control frameworks for AI Agents. By integrating context engineering, architectural constraints, and garbage collection, it transforms unreliable large models into dependable production systems.