Grader-agent
Grader 的 agent 采用五阶段主循环：perceive→plan→act→observe→respond，由 harness 驱动。模型只决定意图与置信度，路由、工具集、风险等级由代码钉死。act 分只读 ReAct、RAG、人工提案、批量初批。四个写动作不在工具表，只能产提案。批改输出结构化 GradingDraft，逐条绑定 rubric 与段落 hash，保证可重放。
erictrainnlpinxihu/Grader-agent
View on GitHub ↗Grader 的 agent 采用五阶段主循环：perceive→plan→act→observe→respond，由 harness 驱动。模型只决定意图与置信度，路由、工具集、风险等级由代码钉死。act 分只读 ReAct、RAG、人工提案、批量初批。四个写动作不在工具表，只能产提案。批改输出结构化 GradingDraft，逐条绑定 rubric 与段落 hash，保证可重放。
Grader-agent ranks 1001st of 1,124 in Domain-Specific Apps, last commit 26 days ago.
Star trajectory
15 days tracked- in 30d, below the Domain-Specific Apps median of +3.0%.
Last commit 26 days ago, staler than the median of 11d.
How it compares in Domain-Specific Apps
Ranked 1001st of 1,124 by stars.
- Language
- Python
- License
- MIT License
- Contributors
- 1
- Forks
- 0
- Watchers
- 0
- Open issues
- 0
- Releases
- 0
- Created
- 2026-09-21
- Stars this week
- -9
You've seen where Grader-agent stands today.
Get the moves before everyone else: the 5 Domain-Specific Apps repos climbing fastest, every Tuesday.
Free weekly digest. Double opt-in, unsubscribe anytime.