GuardianAgentBench (LlamaIndex): leaderboard
GuardianAgentBench (Vectara, 2026) runs tool-using agents through 580 business scenarios in six domains, a third of them with adversarially modified tool responses, and requires both the final answer and the tool-call trace to match the ground truth. Every model runs as a LlamaIndex function-calling agent. Overall correctness in percent averaged over the six domains, judged by Claude Sonnet 4.5; higher is better.
Metric: Overall correctness (%; mean over six agent domains of the share of scenarios whose answer and tool-call trace are both judged correct by Claude Sonnet 4.5; 580 scenarios with 81 tools, a third with adversarially modified tool responses; LlamaIndex FunctionAgent, default API settings). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LlamaIndex + Claude Opus 4.5 | 73.7 | |
| 2 | LlamaIndex + Gemini 3 Pro | 72.3 | |
| 3 | LlamaIndex + GPT-5.2 Pro | 70.8 | |
| 4 | LlamaIndex + Qwen 3 Max | 66.8 | |
| 5 | LlamaIndex + GPT-OSS-120B | 66.1 | |
| 6 | LlamaIndex + DeepSeek V3.2 | 64.2 |
Interactive version: theaggregate.ai/benchmark?slug=guardianagentbench-llamaindex · How It Works · Data refreshed daily, snapshot 2026-11-13.