Source-linked AI summary
Survey on Evaluation of LLM-based Agents
Asaf Yehudai, Lilach Eden, Alan Li, Guy Uziel, Yilun Zhao, Roy Bar-Haim, Arman Cohan, Michal Shmueli-Scheuer
TL;DR
LLM-based agents require evaluation methods that capture sequential decisions and dynamic-environment interaction rather than text generation alone. This survey synthesizes capability, application-specific, generalist-agent, benchmark-dimension, and developer-framework evaluations, finding a shift toward more realistic assessments while identifying unresolved needs for scalable, fine-grained, cost-efficient, safe, and robust evaluation.
Problem
Evaluating adaptive agents requires assessing sequential decision-making, action-based task completion, and changing environments beyond textual outputs.
Method
The survey maps core capabilities, application-specific and generalist benchmarks, benchmark dimensions, evaluation frameworks, and future research directions.
Results
The evaluation field is progressing from isolated capability tests toward realistic, dynamic, challenging environments, while current frontier benchmarks remain difficult.
Takeaways & Limitations
Future evaluation should develop fine-grained, scalable methods and standardized metrics for cost-efficiency, safety, and robustness.
Takeaways & Limitations
Because the field evolves exceptionally rapidly, the survey is inevitably a snapshot and may omit recent or forthcoming developments.
Abstract
from arXiv · showhide
LLM-based agents represent a paradigm shift in AI, enabling autonomous systems to plan, reason, and use tools while interacting with dynamic environments. This paper provides the first comprehensive survey of evaluation methods for these increasingly capable agents. We analyze the field of agent evaluation across five perspectives: (1) Core LLM capabilities needed for agentic workflows, like planning, and tool use; (2) Application-specific benchmarks such as web and SWE agents; (3) Evaluation of generalist agents; (4) Analysis of agent benchmarks' core dimensions; and (5) Evaluation frameworks and tools for agent developers. Our analysis reveals current trends, including a shift toward more realistic, challenging evaluations with continuously updated benchmarks. We also identify critical gaps that future research must address, particularly in assessing cost-efficiency, safety, and robustness, and in developing fine-grained, scalable evaluation methods.
1 Introduction
LLM-based agents extend static language models with multi-step workflows and external tools, enabling interaction with dynamic environments. This shift requires evaluation beyond text outputs, and the survey maps capabilities, applications, benchmarks, and frameworks while identifying research gaps.
- Motivation: LLM-based agents combine a backbone LLM with multi-step workflows and external tools to plan, execute, and adapt in real-world settings.They can perform computations, retrieve current information, and interact with their environment.
- Evaluation Need: Evaluating agents requires measuring sequential decision-making and task completion through action sequences in dynamic environments, not only textual outputs.
- Evaluation Need: Benchmarks must co-evolve with agent capabilities by accommodating new task classes and domains.
- Survey Scope: The survey provides an overview of agent evaluation and maps the landscape for developers, benchmark creators, practitioners, and researchers.
- Survey Scope: It reviews core capabilities, application-specific and generalist-agent benchmarks, benchmark dimensions, evaluation frameworks, and future research directions.
2 Agent Capabilities Evaluation
The survey evaluates four core abilities—planning, tool use, self-reflection, and memory—through isolated capability tests and full agent workflows. Benchmarks increasingly model longer, stateful, real-world interactions, but standardized self-reflection evaluation remains lacking.
- Scope: The survey treats planning, tool use, self-reflection, and memory as core agent abilities evaluable in isolation or within complete workflows.
- Planning and Multi-Step Reasoning: Planning benchmarks range from multi-step reasoning and classical planning to structured workflows and natural-language real-world planning.
- Function Calling & Tool Use: Tool-use evaluation has progressed from synthetic one-step function calls to multi-turn, dependent, constraint-aware, long-context, and real-server interactions.
- Function Calling & Tool Use: Frontier tool-use benchmarks such as MCP Atlas and Tool-Decathlon remain challenging despite significant model advancements.
- Self-Reflection: Self-reflection benchmarks use feedback and prior examples to assess self-correction and decision-making, but no standardized methodology exists.
- Memory: Memory evaluation covers episodic, semantic, and procedural mechanisms, with current methods limited in long-range consistency and dynamic-memory handling.
3 Application-Specific Agents Evaluation
Application-specific evaluation spans web, software engineering, scientific, and conversational agents, moving from simplified tests toward realistic, multimodal, end-to-end environments. These benchmarks increasingly expose long-horizon, interaction, and evaluation-quality limitations.
- Scope: The survey focuses on web, software engineering, scientific, and conversational agents while examining data curation, environments, and metrics.
- Web Agents: Web evaluation has progressed from simplified shopping simulations to realistic offline and online environments supporting rich interactions and intermediate-goal assessment.
- Web Agents: Web benchmarks increasingly test multi-turn dialogue, enterprise workflows, multi-site tasks, granular analysis, semantic judging, and multimodal interaction.
- Web Agents: Online-Mind2Web was proposed as a more rigorous alternative after evidence that WebVoyager may produce over-optimistic performance estimates.
- Software Engineering Agents: SWE-bench evaluates real GitHub issues using repository context, executable environments, and validation tests, extending beyond short algorithmic coding tasks.
- Software Engineering Agents: Model performance on SWE-bench Pro remains below 25% Pass@1, highlighting limitations on long-horizon, multi-file code changes.
4 Generalist Agent Evaluation
Generalist-agent evaluation addresses agents that integrate planning, reasoning, tool use, and interactions across applications and environments. The survey distinguishes broad-capability environments from unified collections of task-specific benchmarks and protocols.
- Motivation: The transition toward general-purpose agents requires integrated capabilities including planning, reasoning, tool use, web interaction, file handling, and code execution.
- Broad-Capability Benchmarks: One evaluation approach uses inherently broad benchmarks such as Gaia, whose tasks require reasoning, multimodality, web browsing, and tool use.
- Broad-Capability Benchmarks: Gaia2 extends required capabilities with ambiguity, noise, temporal constraints, and multi-agent collaboration in a mobile environment.
- Computer Environments: OSWorld, AppWorld, and Gaia2 evaluate complex tasks across applications, differing in whether agents use UI interactions or code and API calls.
- Unified Benchmarking: A second approach unifies task-specific benchmarks across domains, while Harbor and Exgentic pursue unified protocols for cross-environment assessment.
5 Core Benchmark Dimensions
The survey analyzes agent benchmarks across data, environments, interfaces, metrics, and safety, highlighting a shift toward dynamic evaluation and the limits of coarse success measures.
- Benchmarks are analyzed along five dimensions: data curation, environment, interface, metric, and safety.
- Data Curation: Hybrid data curation combines human validation, synthetic worlds with programmatic checks, and cleaned real-world interaction logs.
- Environment: Dynamic environments let agent actions alter observed state, capturing cascading errors and enabling diagnosis of long-horizon failure modes.
- Interaction Interface: Benchmarks use code and terminal, tool-calling, or graphical interfaces to define the agent’s action and observation spaces.
- Metric: Task completion is implemented through application-specific checks including unit tests, gold-state matching, and answer matching, but binary outcomes miss intermediate progress.
- Safety and Robustness: Robustness can be measured with pass^k, while enterprise benchmarks also need tests for privacy, access control, and other safety constraints.
6 Frameworks for Agent Evaluation
Agent-evaluation frameworks extend beyond fixed benchmarks by integrating monitoring, quality assessment, experimentation, and custom scenarios into development workflows. They evaluate final responses, individual steps, and trajectories, while balancing precision against flexibility and leaving important scaling and safety gaps.
- General-purpose frameworks integrate with development processes to support continuous monitoring, evaluation, error analysis, and performance optimization.
- Frameworks assess agent quality at final-response, stepwise, and trajectory levels, with increasingly granular views of execution.
- Final Response Evaluation: Final-response evaluation uses LLM-based judges and customizable metrics but provides limited insight into intermediate decisions, execution efficiency, and failure causes.
- Stepwise Evaluation: Stepwise evaluation localizes errors in generations, tool calls, and routing decisions, but commonly assumes actions can be assessed independently.
- Trajectory-Based Assessment: Trajectory assessment compares observed paths with expected paths or uses reference-free judges to assess coherence, efficiency, and goal directedness.
- Evaluation Method Choices: Reference-based methods offer precision and reproducibility, whereas reference-free methods provide flexibility at lower reliability; judge scope creates a similar trade-off.
- Supporting Capabilities and Challenges: Frameworks support annotation, human-in-the-loop evaluation, synthetic data generation, and A/B comparisons, but struggle with large-scale root-cause analysis and evaluation cost.
7 Discussion
Agent evaluation is moving toward realistic, challenging, and continuously updated benchmarks. The survey identifies remaining needs for granular, efficient, safe, robust, and attribution-aware evaluation.
- Realistic and Challenging Evaluation: Benchmarks increasingly use realistic environments, real GitHub issues, and long-horizon tasks that approximate professional workflows.
- Live Benchmarks: Live benchmarks adapt to faster agent development, correct shortcomings in earlier versions, and accommodate evolving tool ecosystems.
- Advancing Granular Evaluation: Coarse end-to-end success metrics do not diagnose tool selection, reasoning quality, or intermediate failures, motivating standardized step-by-step measures.
- Cost and Efficiency Metrics: Evaluations should track token usage, API expenses, inference time, and resource consumption alongside performance to measure cost efficiency.
- Scaling & Automating: Static human-annotated data limits scalability and becomes outdated, motivating synthetic data and automated evaluation using LLM or agent judges.
- Safety and Compliance: Current benchmarks insufficiently test safety, trustworthiness, policy compliance, adversarial robustness, and bias mitigation, especially in multi-agent scenarios.
- Decoupling LLM & Harness Evaluation: Benchmarks often conflate backbone LLM capabilities with agent-harness design, so controlled protocols are needed to attribute performance gains.
8 Conclusion
The survey traces agent evaluation from isolated capabilities in simplified settings to realistic, dynamic, and challenging environments. It calls for fine-grained, scalable methods and standardized cost-efficiency, safety, and robustness metrics.
- The survey concludes that evaluation has progressed toward realistic, dynamic, and challenging environments but still needs methods beyond overall success rates.
- Future research should develop fine-grained, scalable evaluation and standardized metrics for cost-efficiency, safety, and robustness.
Limitations
The survey offers a structured synthesis of LLM-based agent evaluation, but its coverage is bounded by the field’s rapid evolution, selective scope, and limited depth per benchmark or framework.
- The survey is a snapshot because new benchmarks, frameworks, architectures, and findings appear rapidly, potentially leaving out recent or forthcoming developments.A continuously updated GitHub repository is intended to mitigate this limitation over time.
- Benchmark and framework selection is representative rather than exhaustive, so some niche evaluation approaches receive limited or no detailed coverage.
- Wide topic coverage constrains the depth of analysis for individual benchmarks and frameworks, making primary articles necessary for exhaustive understanding.
- Future-direction and critical-gap discussions involve foresight and interpretation, while research trajectories and the relative importance of identified areas may evolve.
- Despite these limitations, the survey provides a structured synthesis intended as a useful resource for researchers, developers, and practitioners.
A Literature Review Methodology
The survey uses a multi-stage literature review combining systematic search, structured selection, and expert validation to capture the evolving evaluation landscape for LLM-based agents.
- The review comprises three phases: systematic search, structured selection, and expert validation.
- Searches covered Google Scholar, ACL Anthology, HuggingFace Papers, and arXiv using targeted LLM-agent evaluation and benchmarking keywords.
- Forward and backward citation chasing from seminal and highly cited seed papers uncovered foundational work and emerging methods beyond standard keyword indexing.
- Inclusion prioritized novel benchmarks, evaluation frameworks, or significant assessment methodology, while excluding architecture-only work and traditional static LLM evaluation.
- Domain experts, including SWE and web-agent researchers and benchmark creators, reviewed selected manuscripts to validate coverage of state-of-the-art evaluation protocols.
B.1 Planning and Multi-Step Reasoning
Planning, multi-step reasoning, tool use, and self-reflection are evaluated through increasingly specialized benchmarks, but current agents still struggle with long-horizon planning and realistic interactions.
- Planning and Multi-Step Reasoning: Current agents excel at short-term tactical planning but struggle with strategic long-horizon planning across diverse domains.PlanBench provides a dedicated framework for this comparison.
- Planning and Multi-Step Reasoning: Even state-of-the-art agents lag behind classical symbolic planners on everyday planning scenarios.
- Planning and Multi-Step Reasoning: Real-world planning and reasoning require multiple sequential operations, and benchmarks such as Natural Plan show performance decreases as task complexity increases.
- Planning and Multi-Step Reasoning: Planning benchmarks assess task decomposition, state tracking, self-correction, causal understanding, and meta-planning as core abilities.
- Tool Use: Tool-use evaluation has progressed from simple one-step calls with explicit parameters toward multi-turn, multi-step scenarios requiring continuous state management.BFCL v2 and v3 address organizational tools and integrated multi-step evaluation logic, while ToolSandbox adds stateful execution and dynamic milestone evaluation.
- Tool Use: Specialized benchmarks model complex tool interactions through implicit parameter inference, user constraints, long-context processing, and nested API-call sequences.
- Self-Reflection: Self-reflection evaluation uses feedback loops, prior experiences, randomized instructions, and interactive decision-making tasks, but standardized methodology remains a critical gap.
C Application-Specific Agents Evaluation
The survey organizes application-specific agent evaluation around scientific, workplace, web, software engineering, conversational, and general-purpose settings. Its recommendations match benchmark choice to task focus, interaction modality, and environment dynamics, while noting saturation or remaining headroom in several areas.
- Scientific agents: Scientific-agent evaluation spans workflow simulators, virtual discovery environments, and domain-focused suites for biological research.These frameworks measure sequential, end-to-end research workflows, including experimental design and iterative hypothesis refinement.
- Scientific agents: Deep-research agents are evaluated through retrieval quality, knowledge synthesis, and verifiability, with reported performance at 50–70% of human performance.These agents combine search, web, and scientific capabilities to produce citation-backed answers.
- Generalist and workplace agents: Workplace benchmarks simulate enterprise environments requiring browsing, coding, collaboration, policy adherence, UI and API use, and diverse-data integration.TheAgentCompany models a software company, while CRMArena models customer-relationship-management tasks.
- Web agents: For dynamic web interaction, WebArena is the leading option, with top performance reported at 74.3% as of April 19, 2026.The cited result was based on a submission from February 2026, and the survey notes remaining room for improvement.
- Web agents: Mind2Web is the standard for cached-trace offline evaluation, while WebVoyager is recommended for visually grounded online and multimodal agents.Mind2Web-Live and Online-Mind2Web are identified as newer options for updated, reactive environments.
- Software engineering agents: SWE-bench Verified remains the coding-agent gold standard, but approximately 80% top performance suggests near saturation; SWE-bench-pro offers a harder alternative at about 46%.SWE-Lancer targets freelance-style completion, and Terminal Bench targets command-line interactions.
- Conversational and general-purpose agents: τ-bench remains the community standard for dialogue agents requiring user simulation and tool use, despite high current success rates and no widely adopted replacement.For general-purpose agents, GAIA targets reasoning and tool selection, AppWorld structured tool calling, OS-World GUI interaction, and HAL cross-benchmark comparison.