agent-benchmarklisted

# Agent Benchmark Build and run **dynamic, professional benchmarks** for agent systems: skills, tools, prompts, harnesses, and workflows. Popularity is not evidence. Same tasks + measured metrics + reported methodology is evidence. Grounded in field practice from Kun Chen’s agentic engineering workflow talk ([video](https://www.youtube.com/watch?v=iQyg-KypKAA)): viral skills can **increase tokens and worsen outcomes**; GitHub MCP vs CLI/agent-optimized tools differed by **~3× tokens** and **>2× latency** on the **same tasks**; browser tools differed in **turns and tokens** for equal work; long loops only help when a **verifiable metric** exists. See [REFERENCE.md](REFERENCE.md) for principles, metric catalog, and experiment templates. ## Safety - Prefer **read-only or sandboxed** tasks for first runs. Do not hit production write APIs, charge real cards, or delete remote branches/data as part of a bench case unless the user explicitly authorizes a labeled “live” suite. - Never publish secrets, tokens, customer data, or private repo contents in bench artifacts. Redact logs. - Cap cost: set max tokens, max turns, max wall time, and max concurrent trials **before** the first run. - Do not claim statistical significance with n=1. Report n, variance, and limits honestly. - Isolate runs (clean worktree / temp dir / fresh session) so order effects and dirty git state do not poison comparisons. - When benchmarking third-party skills/tools, record exact versions/commits; do not sm