agentbench
exe215 제작
Coding Agents & IDEs요약
Benchmark your OpenClaw agent across 40 real-world tasks.
설명
이 스킬이 하는 일
AgentBench runs 40 real-world tasks across 7 domains to measure how well an OpenClaw agent performs. Tasks cover file creation, research, data analysis, multi-step workflows, memory, and error handling. Results are scored across four weighted layers and can be submitted to a public leaderboard.
Measures actual agent behavior across diverse real-world tasks rather than synthetic or code-only benchmarks, and the objective scoring layers (55% of total weight) can't be self-inflated.
사용 시점
- Checking if a new agent config handles multi-step tasks reliably
- Comparing agent performance before and after changing tool settings
- Running a fast subset of tests to catch regressions after updates
- Identifying which task domains an agent consistently underperforms in
- Submitting scored results to the AgentBench public leaderboard
예시 워크플로
User asks: /benchmark --fast
- 1
Reads task.yaml files from the tasks/ directory, filtering to easy and medium difficulty tasks only
- 2
Creates a timestamped run directory under agentbench-results/
- 3
Executes each task in an isolated workspace, tracking tool calls, errors, and timing
- 4
Scores each task across four layers: structural file checks, metrics analysis, behavioral self-evaluation, and output quality
- 5
Generates results.json with integrity signature, report.md, and report.html in the run directory
Overall score with per-domain breakdown, task detail table, and saved report files ready for review or submission
설치
clawhub install exe215/agentbenchCLI 명령을 실행하거나, 프롬프트를 Claude Code, Cursor, Codex 등 사용하는 코딩 에이전트에 붙여넣으세요.