← 开源
stanford-iris-lab

meta-harness-tbench2-artifact

Meta-Harness: 76.4% on Terminal-Bench 2.0 (Claude Opus 4.6)

AI EngineeringTest & guardBuild agent workflow (SDK)Python
在 GitHub 打开
增长势头
+-124 小时新增 Star-0.1%
1.23k
Star
169
Fork
+5
本周
1
贡献者
创建于 2026-03-26 · 更新于 2026-10-16 · 今日第 18912 名
主要开发者
yoonholee
README

Meta-Harness

Agent scaffold for Terminal-Bench 2.0, built on top of Terminus-KIRA by KRAFTON AI and Harbor's Terminus-2 framework.

Results

76.4% on Terminal-Bench 2.0 (89 tasks × 5 trials, Claude Opus 4.6).

Split N Score
Easy 4 100.0
Medium 55 81.1
Hard 30 64.7
All 89 76.4

Usage

pip install harbor

export ANTHROPIC_API_KEY=

harbor run \
  --agent-import-path agent:AgentHarness \
  -d terminal-bench@2.0 \
  -m anthropic/claude-opus-4-6 \
  -e runloop \
  -n 20 \
  --n-attempts 5

Method

Meta-Harness extends the Terminus-KIRA agent with environment bootstrapping: before the agent loop starts, it gathers a snapshot of the sandbox environment (working directory, file listing, available languages/tools, package managers, memory) and injects it into the initial prompt. This saves 2-5 early exploration turns that the agent normally spends on ls, which python3, etc.

The agent was discovered through automated harness evolution. More details coming soon.

Acknowledgements

We thank KRAFTON AI for compute support.

排名相邻
同类项目