AgentEvalBench (Agent-Sourcecode): leaderboard
Automated agent evaluation on AgentEvalBench (Zhou et al., 2026, AWS AI Labs): for 20 real LLM agents with pre-collected execution traces, each under a generic and a specific requirement, the system writes evaluation code for the agent as a Claude Code agent (up to 50 tool calls) that sees the agent source code but no traces. Eval@1 in percent over the 40 cases, counting code that runs and produces non-vacuous results on the real traces at the first attempt; code that runs on mocked data or returns constant metrics fails. Higher is better.
Metric: Eval@1 (%; share of 40 cases, 20 agents under generic and specific requirements, whose generated evaluation code runs and yields non-vacuous results on real agent traces at the first attempt; Claude Code v2.1.92 agent with the source code only, up to 50 tool calls). Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 60 | #244 |
| 2 | Claude Haiku 4.5 | 45 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentevalbench-agent-sourcecode · How It Works · Data refreshed daily, snapshot 2026-11-12.