The Benchmark — AgentBench

AgentBench

Evaluates how well language models can act as autonomous agents by completing multi-step tasks in simulated digital environments without human intervention.

What it measures

AgentBench tests whether models can operate independently in realistic interactive scenarios—browsing the web, using APIs, writing and running code, interacting with databases—rather than just answering isolated questions. It measures planning, tool use, error recovery, and the ability to decompose complex goals into executable actions. Tasks range from booking flights to debugging code to querying databases, all within controlled simulated environments.

Why it was created

By late 2023, most AI benchmarks evaluated models on static, well-formatted tasks. But deployment reality showed that models often fail when they need to act in open-ended environments: deciding which tool to call, handling API failures, adapting when plans don’t work. AgentBench (released by Tsinghua University, Alipay, and others in April 2023) was designed to stress-test this gap—whether models could autonomously solve problems that require exploration and iteration, not just knowledge recall.

How it works

AgentBench contains 8 domains with 1,200+ tasks total. Agents operate in sandboxed environments: web browsing (WebShop), code execution (HotpotQA programmatic variant), database queries, and Hugging Face API interactions. A model receives a natural language goal (“Book me a flight from NYC to LA on March 15 for under $400”), then generates actions iteratively. Each action gets a response from the environment. Scoring is binary (task completed successfully) or scalar (progress toward goal). The benchmark measures success rate and average number of steps taken.

What scores mean in practice

GPT-4 achieves ~55% success on WebShop and ~60-65% on database/API tasks. GPT-3.5 sits around 25-35%. Claude 2 scores ~45-50%. A year ago (2023), passing 40% on most AgentBench domains would have been competitive; today that’s a weak baseline. Notably, there’s a huge gap between “knows how to use a tool” (captured by HumanEval) and “autonomously orchestrates multi-tool workflows”—models that score 85% on HumanEval often score 50-60% here. Human performance varies by task; for WebShop, humans achieve ~90% with realistic time constraints.

Known limitations

• Environment brittleness and reward engineering: Success rates depend heavily on exact environment simulation. WebShop’s HTML is somewhat simplified; real web interfaces are messier. Minor changes in how the simulated environment returns errors can significantly shift scores, raising questions about real-world transfer.

• Small task diversity within domains: AgentBench has fewer than 200 unique WebShop tasks, raising concerns about data contamination and overfitting. As models get trained on benchmark data, gains may not reflect genuine improvement in agent capabilities.

• Single-agent, no adversarial scenarios: Tasks assume a cooperative environment with well-formed APIs. Real-world agents face rate limiting, ambiguous error messages, adversarial inputs, and incomplete documentation—none of which AgentBench systematically tests.

When to trust it (and when not to)

• Trust it for: Comparing models’ ability to chain tool calls and iterate on failures. If one model scores significantly higher than another on AgentBench, it’s likely genuinely better at planning and error recovery in constrained interactive settings. Useful for distinguishing models that have tool-use scaffolding vs. those that don’t.

• Don’t trust it for: Predicting real-world agent performance. A model’s AgentBench score tells you almost nothing about whether it will handle production APIs, malformed responses, or security constraints. The benchmark’s simplified environments hide the brittleness problems that dominate actual deployment. Use it as a component of evaluation, not a standalone signal.