Sample record. This work is fictional demonstration content. Results, people and institutions are placeholders; references in full-text articles are real publications.
Highlights
- A four-axis framework for describing agent evaluations.
- Common weaknesses: unvalidated model judges, single runs, unreported cost.
- A reporting checklist for agent evaluations.
Abstract
As language-model agents move from demonstrations to deployments, the way they are evaluated determines which claims can be trusted. We review how agents are evaluated in the literature, organising practice along four axes: the task environment, the success criterion, the treatment of variance, and the reporting of cost. We identify recurring weaknesses: success is often judged by another model without validation, runs are rarely repeated, cost and latency are reported in under a third of studies, and environment versions are seldom pinned. We close with a checklist for evaluation reports and open problems, including measurement of safety-relevant side effects.
Keywords evaluation · agents · benchmarks · reproducibility · survey
1.Introduction
Language models are increasingly deployed as agents that plan, call tools and act over many steps [1, 2]. Evaluating such systems is harder than evaluating a classifier: tasks are long, outcomes depend on external state, and the same agent can behave differently across runs. Benchmarks for agents have multiplied [3], and so has the number of ways to report their results.
This survey asks what a reader can conclude from a published agent evaluation. We do not rank systems; we analyse the evaluation designs themselves, building on earlier critiques of benchmarking practice [4, 5].
2.A four-axis framework
We describe each evaluation by four properties.
- Environment — whether tasks run in a simulator, a sandboxed real system or a static trace.
- Success criterion — programmatic check, human judgement or model-based judgement.
- Variance — the number of repeated runs, seeds and task samples, and how uncertainty is reported.
- Cost — tokens, tool calls, wall-clock time and money per task.
Table 1 summarises the prevalence of each choice among the 64 evaluations we coded.
| Axis | Choice | Studies | Share |
|---|---|---|---|
| Environment | Simulator | 29 | 45% |
| Environment | Sandboxed real system | 21 | 33% |
| Environment | Static trace | 14 | 22% |
| Success | Programmatic check | 27 | 42% |
| Success | Model-based judge | 26 | 41% |
| Success | Human judgement | 11 | 17% |
| Variance | Repeated runs reported | 12 | 19% |
| Cost | Cost or latency reported | 18 | 28% |
3.Findings
Success criteria are weakly validated
Model-based judging is now the most common criterion for open-ended tasks, but only a minority of studies report agreement with human raters. Where agreement is reported it is often measured on a small sample and not stratified by task difficulty.
Variance is under-reported
Most studies run each task once. Where variance has been measured, differences between seeds can be as large as the differences between methods [6]. Holistic benchmarks for language models [7] provide a model for reporting multiple metrics, but equivalent norms for agents are still emerging.
Cost is an afterthought
Fewer than one in three studies report token or tool-call cost, even though cost often differs by an order of magnitude between agents with similar success rates. Without it, a more accurate agent cannot be compared fairly with a cheaper one.
4.A reporting checklist
- Pin and publish versions of environments, tools and prompts.
- Report the success criterion and, for model judges, agreement with humans on a stratified sample.
- Repeat runs; report mean and an interval, with the number of seeds.
- Report tokens, tool calls, wall-clock time and cost per task.
- Report failures by category, not only the success rate.
- State which side effects were monitored (for example unintended writes).
5.Open problems
Open problems include evaluating safety-relevant side effects that do not affect task success, benchmarks that stay valid as agents improve, and methods for comparing agents under equal budgets. Foundation-model risks more generally are discussed in [8].
Declarations
- Author contributions
- K.M. and I.N. coded the studies and wrote the first draft; L.H. contributed to the framework and revised the manuscript.
- Funding
- Sample article – no funding to declare.
- Competing interests
- The authors declare no competing interests.
- Data and code availability
- The coding sheet would be deposited in a public repository. Sample article. All results, names and affiliations in this demonstration article are synthetic placeholders.
References
- [1]Wang, L., Ma, C., Feng, X., Zhang, Z., et al. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6), 186345. arXiv:2308.11432
- [2]Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. International Conference on Learning Representations. arXiv:2210.03629
- [3]Liu, X., Yu, H., Zhang, H., Xu, Y., et al. (2024). AgentBench: Evaluating LLMs as agents. International Conference on Learning Representations. arXiv:2308.03688
- [4]Raji, I. D., Bender, E. M., Paullada, A., Denton, E., Hanna, A. (2021). AI and the everything in the whole wide world benchmark. NeurIPS Datasets and Benchmarks Track. arXiv:2111.15366
- [5]Bowman, S. R., Dahl, G. E. (2021). What will it take to fix benchmarking in natural language understanding?. Proceedings of NAACL-HLT, 4843–4855. arXiv:2104.02145
- [6]Bouthillier, X., Delaunay, P., Bronzi, M., Trofimov, A., et al. (2021). Accounting for variance in machine learning benchmarks. Proceedings of Machine Learning and Systems, 3. arXiv:2103.03098
- [7]Liang, P., Bommasani, R., Lee, T., Tsipras, D., et al. (2023). Holistic evaluation of language models. Transactions on Machine Learning Research. arXiv:2211.09110
- [8]Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., et al. (2021). On the opportunities and risks of foundation models. arXiv preprint. arXiv:2108.07258
Peer review
Open reports from independent reviewers. See how outcomes are determined.
- Status
- Peer reviewed
- Requested
- 18 Dec 2025
- Completed
- 28 Apr 2026
- Report identity
- Signed reports
Reports on version 1
- Approved with reservations
Dr. Pavel Novotný
A timely survey with a useful four-axis framework. The selection criteria for included papers need to be more explicit so that the review can be repeated.
Strengths
- The framework organises a scattered literature.
- The reporting checklist is directly usable.
Concerns
- Search strategy and inclusion criteria are described only briefly.
- Some benchmarks are discussed from secondary sources.
Authors' response
We added a PRISMA-style flow diagram, the full search strings and the screening protocol, and re-read the primary sources.
- Approved with reservations
Dr. Ioana Pop
Valuable synthesis. I would like the claims about how often variance is reported to be backed by counts.
Strengths
- Strong coverage of recent work.
- Balanced tone.
Concerns
- The statement that 'most studies' ignore variance needs numbers.
Authors' response
Section 4 now reports counts for 112 papers.
Reports on version 2Current version
- Approved
Dr. Pavel Novotný
The revised search protocol makes the survey reproducible. Approved.
Strengths
- Transparent methodology.
Concerns
- None.
- Approved
Dr. Ioana Pop
My concerns are resolved, and the counts make the central argument much stronger.
Strengths
- Evidence now supports each claim.
Concerns
- None.
Versions
Version 2 · 22 May 2026Latest
Revised after review: search protocol and counts added.
10.xxxx/ctm.pb.2026.0003.v2
Version 1 · 15 December 2025
Posted; peer review requested.
10.xxxx/ctm.pb.2026.0003.v1
How to cite
Mensah, K., Novak, I., & Hartwell, L. (2026). A Survey of Evaluation Practices for Language-Model Agents. CONTINUUM Publishing, CTM-PB-2026-0003. doi:10.xxxx/ctm.pb.2026.0003
© 2026 The authors. Published by AEX CONTINUUM under the CC BY 4.0 licence.