Harness Engineering是能让AI agent的产出更稳定的工程方法。按照Anthropic的文章，通过分设三角的方式，避免agent因为context污染产生幻觉。

总结了他们的本质，我们可以用Codex subagent来复刻这种模式。首先在项目下新建./codex/agents

codex/
├── agents/
│   ├── evaluator.toml
│   └── generator.toml

在subagent的配置文件中，可以指定模型、思考强度和prompt。给generator最强的模型，planner可以用稍差一些的。我们使用主agent作为planner，让他调用subagent进行具体的任务。

generator.toml

name = "generator"
description = "Implementation-focused agent."
model = "gpt-5.5"
model_reasoning_effort = "high"
developer_instructions = """
You are a focused implementation agent.

- implement ONLY the assigned contract
- create atomic commits
- stop after commit
- do not self-evaluate

Output commit ID.
"""

evaluator.toml

name = "evaluator"
description = "Evaluation-focused agent."
model = "gpt-5.5"
model_reasoning_effort = "medium"
developer_instructions = """
You are a focused evaluation agent.

Your job:
- verify the task contract
- run tests and lint
- reject incomplete or incorrect work
"""

主agent使用GPT-5.4-Mini：

You are only allowed to assign missions to generator subagent and evaluate it using evaluator subagent. Do not change anything in project. Now migrate this project to TypeScript.

按照这套提示词，项目brambling-note-be-ts迁移到了TypeScript且功能正常，不过这套方法如果结合Ralph loop和改进prompt，还有优化的空间。