Meta-Harness: What to call the system that calls Claude Code and Codex
The system that calls Claude Code, Codex, Cursor, or Cline is a meta-harness.
Claude Code and Codex are agent harnesses: runtimes that let a model act through tools. More people are building systems above those runtimes to choose an agent, supply the task and context, check the result, and send the work elsewhere for review. Calling both layers a "harness" makes it hard to explain what changed or compare results.
Meta-harness is the most useful name I have found for the outer layer. Anthropic uses the term for Managed Agents, which can run different agent harnesses without depending on any one of them.
meta-harness
-> agent harness
-> Claude Code / Codex
-> Cursor / Cline
-> model + tools + skills
<- trajectory + outcome
| Term | Meaning |
|---|---|
| Meta-harness | The outer system that plans, routes, verifies, and judges work across agent harnesses. |
| Agent harness | The runtime that lets a model act through tools, such as Claude Code or Codex. |
| Run configuration | The pinned model, harness, effort level, skills, tools, and settings used for a run. |
| Trial | One attempt at a task. |
| Trajectory | The recorded path through that attempt: messages, tool calls, intermediate results, and output. |
A skill that improves one model-harness configuration can make another worse. You may need to rewrite the skill for a different model or harness, then inspect the trajectory to see why its behavior changed.