A Survey on Evaluation of LLM-based Agents

core: agent evaluation spans four axes — core capabilities (planning, function calling, self-reflection, memory), application-specific benchmarks (web, SWE, scientific, conversational), generalist-agent platforms (single-task unification, full computer environments, standardized protocols), and framework tools (LangSmith, Vertex AI, Patronus AI, Galileo, Mosaic AI). A five-axis structural lens — data curation · environment · interface · metric · safety — is applied to 8 representative benchmarks via a structural sheet, with realistic + live trends plus five future gaps named in the closing section (granularity, cost, scaling, safety, harness decoupling).

Five panes: core capabilities taxonomy, application-specific benchmarks, generalist-agent platforms, five-dimension structural comparison, framework feature matrix, and current trends plus future gaps.

primer — what you need before you start

the task. Yehudai 2025 is a survey of LLM agent evaluation — 30+ benchmarks, 4 capability groups, 4 application domains, 5 structural dimensions. Goal: a structural lens so you can compare benchmarks apples-to-apples instead of relying on the headline number.

how it was done before. Agent benchmarks used to report a single accuracy number on a single task. Different papers used different metrics (Pass@1, Win Rate, exact-match) on different domains. The survey proposes a 5-axis structural lens (Data · Environment · Interface · Metric · Safety) so you can see what a benchmark actually tests.

glossary — terms you will see.
benchmark — a fixed task with a fixed metric; lets you compare models.
agent — LM + controller + tool access; multi-step task completion.
function calling — the agent emits a structured call to a tool, gets a result back.
self-reflection — the agent evaluates its own plan / progress / failure.
long-horizon task — many sequential steps; weeks of human time compressed into hours.
SWE-bench — real GitHub issues; benchmark for code agents.
WebArena — real web tasks; benchmark for browsing agents.
τ-bench — customer-service benchmark with real DBs.

symbol table — what the letters mean.
$D$ — the dataset / task distribution.
$E$ — the environment (real DB, mock web, REPL, etc.).
$I$ — the interface (text input, function-call spec, GUI, etc.).
$M$ — the metric (Pass@1, Win Rate, factuality score, etc.).
$S$ — the safety dimension (what failure modes are measured, what guardrails).
$\pi$ — the policy under test (the agent).

tone legend — what the [brackets] mean.
[obs] observation   [cap] capability group
[app] application domain   [dim] structural dimension
[eq] equation   [take] key takeaway
[sys] system / meta   [limit] paper-reported limit

Six panes cover, in pedagogical order: core capabilities taxonomy; application-specific benchmarks; generalist-agent platforms; five-dimension structural comparison; framework feature matrix + trends; future gaps. Each pane ends with key takeaways so you can stop after any section and still know what you learned.

point: the survey groups agent benchmarks by what they measure — Planning & Multi-Step Reasoning (PlanBench, FlowBench), Function Calling & Tool Use (BFCL, ToolBench, ComplexFuncBench), Self-Reflection (LLM-Evolve, LLF-Bench), and Memory (StreamBench, MemoryAgentBench). Each capability group ships a small number of representative benchmarks; cross-capability synthesis is rare.

Four capability trees printed sequentially — each ASCII tree is the structural claim. Benchmark names and first-author citations are verbatim.

capability representative benchmark system

point: Web Agents (WebShop, Mind2Web, WebArena) test GUI / web interaction; Software Engineering Agents (SWE-bench Verified with 500 samples, SWE-bench Pro with 1,865 tasks, SWE-Lancer with $1M payouts) test code-edit workflows; Scientific Agents (ScienceAgentBench, CORE-Bench, PaperBench) test research workflows; Conversational Agents (τ-Bench, τ²-Bench in Telecom) test multi-turn stateful dialogue.

Four domain trees printed sequentially — the topology is the claim; only the labels change. Benchmark names, sample counts, and dollar payouts are verbatim.

application domain representative benchmark system

point: Single-task unification hosts many task types under one harness — AgentBench (OS + DB + web + game), HAL (Holistic Agent Leaderboard), TheAgentCompany. Full computer environments hand the agent an OS or app — OSWorld, AppWorld, Gaia2 (mobile). Standardized protocols (Harbor Framework, Exgentic) decouple the harness from the model so cross-platform scores become comparable.

Three category trees printed sequentially — the topology is the claim. Platform names and citations are verbatim.

platform category representative platform system

point: the survey's central structural contribution — a five-dimension structural sheet places 8 representative benchmarks along five axes: Data · Env. · Interface · Metric · Safety. Data is mostly Hybrid (curated + LLM-generated); only GAIA is Human-curated. Environment is mostly Dynamic; only Mind2Web is Static. Interface splits by domain — Code for SWE, GUI for web, Tools for full-env, Mix for GAIA. Metric varies — Unit Tests / End-to-end / Action Match / State Match / Answer Match are not directly comparable across rows. Safety is sparsely populated — only TXU-Bench (of 8) reports Safety = Yes.

Verbatim five-dimension structural sheet — 8 benchmark rows × 5 dimension columns. All 40 cells are verbatim; only 8 representative entries are shown.

dimension verbatim cell interpretation system

point: the framework feature matrix lays out 8 framework tools across 6 capability columns (Stepwise / Monitoring / Trajectory / Human-in-Loop / Synthetic Data / A/B). LangSmith and Patronus AI cover all 6 columns; AgentEvals covers only Human-in-Loop. The closing section names two current trends — Realistic and Challenging Evaluation (move toward real environments: WebArena, SWE-bench Verified) and Live Benchmarks (continuous updates to prevent contamination: BFCL v2/v3, SWE-bench Verified → SWE-bench Pro). Five future gaps remain open: granularity, cost / efficiency, scaling, safety, harness decoupling.

Framework feature matrix sheet printed once; trend names and future-gap names are verbatim.

framework capability column future gap system