agent-benchmarks 开源项目推荐（18 个）

ktwu01 /benchmark-radar

Track 42,331+ AI benchmark, eval, dataset, and data-quality records from 40 public sources, with linked evidence and daily updates.

KaiWU5 /Awesome-AI4AI

Updated weekly — until humanity achieves AI4AI. 1148 curated papers across Benchmarks, Harness Design, and Model Design — from long-horizon agents to recursive self-improvement. Citation rankings, GitHub stars, and source-verified weekly news.

REAL-Lab-NU /Awesome-OpenClaw-Papers

这是我们调查《OpenClaw 生态系统调查：从平台可扩展性到约束设计》的官方配套仓库——这是一份为 OpenClaw AI 智能体生态系统策划的论文、基准测试、安全报告、数据集和工具合集。

Andromede-AI /messier

[EMNLP 2026] Messier is a cross-benchmark corpus of AI agent evaluations, linking tasks and results with models, scaffolds, verifiers, scoring rules, and execution trajectories.

Mazha0309 /the-evil-repository

An evidence-hostile, container-isolated benchmark and behavior analysis platform for long-horizon AI software agents.

1690940255ran-dot /benchshield

Anti-cheating audit and isolated evaluation sandbox for AI agent benchmarks: detect the 7 vulnerability patterns, grade on the Agent-Eval Checklist, red-team with zero-capability agents. Zero dependencies.

reacher-z /AI-Can-What

Source-backed directory of AI-built software, agent environments, benchmarks and datasets, with original research and bilingual pages.

jatinsihag2345 /awesome-ai-evaluations

A curated list of frontier AI evaluation benchmarks, long-horizon agent test suites, terminal sandboxes, RLVR math verifiers, and red-teaming datasets.

Alpha-Park /genpark-tool-call-argument-accuracy-validator-skill

Strict type, range, and semantic key validator for agent tool call invocations

kenantang /iclr26-agent-cost

Generated by Codex. Read with caution.

agenticluke /coding-agent-bench-plus

A free Claude Code skill to test coding agents on the same tasks; compare pass rate, cost, and time—built on affaan-m/ECC.

Alpha-Park /genpark-agent-trajectory-pass-fail-evaluator-skill

Step-by-step agent trajectory evaluator comparing execution traces against golden tool call sequences

Numan5837 /Numan5837

Mercor Real World Engineer | Terminal-Bench Domain Expert | Agent evaluation, verifier engineering & reproducible systems

bsha6 /agent-release-gate

Deterministic release decisions from agent benchmark evidence.

CODER7657 /fix-broken-Terminal-Bench-2-Harbor-

A broken Terminal-Bench 2 / Harbor task taken apart: wrong artifact path, unpinned image, leaked solution and a gameable verifier, each found and fixed.

ManavPratapSingh /AgentBench-Lite

Deterministic evaluation harness for multi-turn LLM tool-use agents with hermetic in-memory SQLite sandboxing, automated failure taxonomy, fault injection, and zero-cost replay.

L-X-J /tank-battle-agent-benchmark

用同一套提示词评测多模型 Agent 从零实现完整可玩游戏的能力：固定技术栈、固定评分标准、按模型归档对比。| A standardized prompt for benchmarking LLM agents on building a complete game from scratch — fixed tech stack, fixed rubric, per-model results.

ZouhairMudakka /agent-expirycheck

Inventory expiry regression tests for AI agent business simulators. Detect stock-age resets and expired sales in pinned Vending Bench and RetailBench implementations; no LLM calls.