📝Harness Engineering is the practice of designing the software that wraps a language model into an agent: its tools, instructions, context loading, permissions, and control loop.
A model on its own only maps text to text. The 📝agent harness around it decides what the model sees on each turn, which tools it can call, how results are fed back, and when the loop stops. Examples include 📝Claude Code, Codex, and custom harnesses built directly on model APIs. Two harnesses running the same model can produce very different results.
The work spans several layers. Instruction files such as 📝CLAUDE.md set standing rules; 📝Hooks enforce rules deterministically at lifecycle events; 📝Skill files and 📝MCP Server connections add capabilities on demand; permission modes and sandboxes bound what the agent may do; and compaction and memory policies manage the 📝context window across long sessions.
Harness engineering sits alongside context engineering and memory engineering, which focus on what information the model receives and what persists between sessions; 📝The Difference Between Context Engineering, Memory Engineering, and Harness Engineering separates the three. As frontier models converge in capability, practitioners increasingly point to the harness as the main source of difference in agent performance.