agent-readings · AgentBench: Evaluating LLMs as Agents

Why It Matters

For agent infrastructure, AgentBench is an early systems-shaped benchmark: it needs task servers, isolated workers, interaction histories, score normalization, and environment-specific validators.

Problem

Traditional NLP benchmarks evaluate static input-output behavior rather than multi-turn action in environments.
Many early agent evaluations focused on a single environment, which hid cross-domain weaknesses.
Embodied or multimodal simulators were not always practical for evaluating text-only LLM agents.
Agent failure modes such as invalid actions, format errors, and turn-limit exhaustion were not consistently tracked across tasks.

Key gap: The paper asks how to systematically evaluate LLMs as autonomous agents across multiple realistic interactive task families.

Method

Wrap eight heterogeneous environments in a common multi-turn interface, prompt LLMs with thought/action style outputs, execute environment transitions through task workers, and report both per-environment and normalized overall scores.

Define eight environments grouped as code-grounded, game-grounded, and web-grounded tasks.
Adapt or build each environment so a text-only LLM can interact in multi-turn dialogue/action form.
Use Chain-of-Thought-style thought/action prompting with greedy decoding for reproducible baseline comparisons.
Run API and open-source LLMs through an HTTP/server-client evaluation toolkit.
Host complex task environments in Docker images and separate task workers to avoid environment conflicts.
Score each environment with its task-specific metric and compute a weighted overall score.
Analyze model gaps and episode failure reasons across environments.

Claims

C1 · paper-supported

AgentBench defines eight distinct interactive environments for evaluating LLMs as agents.

Scope: Abstract, Introduction, and Section 3.

E-eight-envs: AgentBench is introduced as a benchmark with 8 distinct interactive environments for assessing LLM-as-Agent reasoning and decision-making.

C2 · paper-supported

The eight environments are grouped into code-grounded, game-grounded, and web-grounded task families.

Scope: Section 3.

E-env-taxonomy: The environments are Operating System, Database, Knowledge Graph, Digital Card Game, Lateral Thinking Puzzles, House-Holding, Web Shopping, and Web Browsing, grouped as code, game, and web.

C3 · paper-supported

AgentBench evaluates multi-turn, open-ended agent behavior and reports environment-specific metrics plus a weighted overall score.

Scope: Section 4.1.

E-metrics: The paper reports per-environment metrics including success rate, F1, reward, game progress, and step success, then computes a weighted overall score to reduce dominance by easier tasks.

C4 · paper-supported

The paper evaluates 27 API-based and open-source LLMs and finds a large gap between top commercial models and many OSS models at the time.

Scope: Abstract, Introduction, and Table 3.

E-model-results: The paper evaluates 27 API-based and OSS LLMs and reports GPT-4 leading the weighted overall score, with a clear gap to OSS models included in the study.

C5 · paper-supported

The paper identifies poor long-term reasoning, decision-making, and instruction following as major obstacles for usable LLM agents.

Scope: Abstract and Section 2.

E-failure-reasons: The authors identify long-term reasoning, decision-making, and instruction following as obstacles, with invalid format/action and task-limit exceeded as common diagnostics.

C6 · paper-supported

AgentBench ships an integrated evaluation toolkit with server-client model access, Dockerized environments, and isolated task workers.

Scope: Introduction, Section 4.1, and repository README.

E-toolkit: AgentBench provides datasets, environments, and an evaluation package using HTTP model access, Docker images, task workers, and a controller/assigner-style workflow.

C7 · paper-supported

AgentBench explicitly tracks finish reasons such as invalid format, invalid action, context-limit exceeded, task-limit exceeded, and complete.

Scope: Section 2.

E-finish-reasons: The paper lists context limit exceeded, invalid format, invalid action, task limit exceeded, and complete as typical finish reasons for agent episodes.

Agent Infrastructure Takeaways

A serious agent benchmark needs a harness, not just prompts: task servers, workers, model adapters, logs, and environment reset all matter.
Cross-domain evaluation is useful because code, game, and web tasks stress different mixes of planning, state tracking, action validity, and domain knowledge.
Tracking finish reasons is a practical debugging primitive for agent systems and should be preserved in traces.
Dockerized task workers are an early example of benchmark-side isolation for heterogeneous agent environments.
Weighted aggregate scores are convenient for leaderboards but should not hide per-environment collapse.
AgentBench is a bridge from ReAct-style text agents toward later executable benchmarks such as WebArena, tau-bench, and OSWorld.

Caveats

The headline model results are historical and should not be treated as current model capability.
The benchmark is broad, but some environments are adapted or simplified compared with live production systems.
Text-only interaction makes the benchmark cheaper and broader, but misses multimodal GUI grounding that OSWorld later targets.
The repository has evolved beyond the original paper version, including function-calling and newer benchmark variants.
Weighted overall scores can obscure which environment caused an agent to fail.
Some task environments have nontrivial setup and resource requirements, including Docker images and service dependencies.