Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Build the web for agents, not agents for the web

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that web agents should not be forced to navigate human-designed browser UIs and web APIs; instead, the research community should build standardized Agentic Web Interfaces (AWIs) designed for agent use, and that such…

desk verdict A clear, well-scoped position paper arguing for agent-native web interfaces; the 'essential' claim is an argument, not a result, but the paper is honest about that and worth engaging. read the letter →

arxiv 2506.10953 v1 pith:34MQPQNV submitted 2025-06-12 cs.LG cs.CL

Xing Han Lù , Gaurav Kamath , Marius Mosbach , Siva Reddy This is my paper
classification cs.LGcs.CL
keywords agenticwebinterfaceagentspositionpaperdesignLLMautomationAIsafetybrowser
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the main obstacle to capable, safe, and affordable web agents is not the agents themselves but the interfaces they are forced to use. Human-oriented browser UIs drown agents in huge DOM trees or token-inefficient screenshots, while developer-oriented web APIs offer too narrow an action space and expose unsafe bypasses. The paper therefore proposes a new kind of artifact, an Agentic Web Interface (AWI), built specifically for agents, and claims that designing such interfaces is essential for web agent performance and would directly fix many current problems. It offers six guiding principles (standardized, human-centric, safe, optimally represented, efficient to host, developer-friendly) and five concrete design suggestions, while explicitly providing no prototype or implementation. A sympathetic reader would take the paper as a call to reorient the field from building agents that navigate human interfaces to jointly building interfaces that agents can natively use.

What carries the argument

The central object is the Agentic Web Interface (AWI), defined as an interface specifically designed for agents to navigate a website, analogous to how a user interface is designed for humans. The argument is carried by two complementary design layers. The first is a set of six guiding principles: standardized, human-centric, safe, optimal representations, efficient to host, and developer-friendly. The second is five concrete suggestions: unified higher-level actions that compose primitive operations, compatibility with human user interfaces through bidirectional translation tools, access control lists plus privacy-preserving password managers for agents, progressive information transfer that sends resized or embedded content, and agentic task queues that cap concurrent agent traffic to protect human users. Together these mechanisms are what would let AWIs fix the representational, resource, and safety issues the paper attributes to current human-oriented interfaces.

What would settle it

Build an AWI for an existing website and compare agents using it against state-of-the-art DOM/screenshot/API-based agents on the same task benchmark with fixed success criteria, token cost, and safety violation counts. If the AWI-guided agent does not improve on at least one of these metrics while matching the others, the paper's claim that AWIs are essential for performance and directly fix current problems would be contradicted. A separate test would be whether any real website operator adopts an AWI voluntarily after being offered the design; if none do despite demonstrated benefits, the proposal's central adoption premise fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that web agents should not be forced to adapt to interfaces designed for humans, neither the visual browser UI nor developer-facing web APIs, and that the research community should instead design Agentic Web Interfaces: standardized interfaces whose structure, action space, and state representations are optimized for agent use. The paper argues this is essential for web agent performance and will directly resolve representational inefficiency, resource strain on websites, narrow action spaces, and several safety and privacy risks. The argument is supported by a diagnosis of current approaches: screenshots hide occluded elements, DOM trees can exceed a million tokens, web APIs cannot perform stateful actions like sorting a list after viewing it, and internal APIs can bypass guardrails such as password prompts. The paper concludes that building AWIs should be a collaborative, iterative, multi-stakeholder effort, with machine learning researchers involved from the start.

Load-bearing premise

The proposal depends on website owners, developers, and standards bodies voluntarily adopting, implementing, and maintaining AWI interfaces, yet the paper provides neither a working prototype nor evidence of economic or political incentives for them to do so.

Editorial extensions

If this is right

  • A standardized AWI action space would let the same high-level actions work across websites, so policy learning and generalization no longer depend on each site's DOM structure.
  • Optimal state representations would cut token costs; the paper estimates current DOM-based interaction with a large model can cost roughly $40 for a single 20-step task, a cost AWIs would reduce.
  • Built-in access control and privacy-preserving password management would let agents use sensitive information without exposing full credentials, addressing unauthorized-purchase and harmful-message risks.
  • Agentic task queues would cap concurrency and spread load, keeping websites responsive for humans while absorbing growing agent traffic.
  • Stateful AWI interactions would enable actions that depend on previous states, such as sorting an already-retrieved product list, without resending the entire query as stateless protocols require.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: If AWIs become widespread, web-agent benchmarks would likely move from measuring how well agents cope with human UIs to measuring task competence on agent-native interfaces, which would change what 'generalization across websites' means and make many current leaderboards non-comparable.
  • Beyond the paper: The paper implicitly predicts an adoption economy, namely that websites will implement AWIs because task queues and efficient representations reduce hosting and bandwidth costs, but this is a testable empirical claim, not yet a demonstrated one; a pilot deployment on a high-traffic site could falsify or support it.
  • Beyond the paper: Progressive information transfer and embedding-based representations shift content processing to website infrastructure, so one unaddressed design question is who decides what an agent may see and retain, and under what consent terms.
  • Beyond the paper: Because the paper distinguishes AWIs from protocols like MCP while allowing them to be combined, an open design question is whether a future AWI standard will simply become a protocol with a stateful layer, blurring the interface/protocol boundary.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This position paper argues that the current entry points for web agents—browser UIs (DOM trees, screenshots, accessibility trees) and web APIs—are designed for humans and are fundamentally mismatched with LLM capabilities. The authors introduce the concept of an Agentic Web Interface (AWI), a unified interface designed specifically for agents, and claim that AWIs are "essential for web agent performance" (Section 3). They list six guiding principles (standardized, human-centric, safe, optimal representations, efficient to host, developer-friendly) and offer concrete suggestions such as unified higher-level actions, access control lists, progressive information transfer, and agentic task queues (Sections 3.2 and 3.3). The paper also argues that the broader ML community must be involved in AWI design (Section 4) and distinguishes AWIs from the Model Context Protocol (MCP) (Section 5). The manuscript explicitly states that no prototype or implementation is provided.

Significance. If the core claim were established, AWIs could be an important reorientation for web-agent research, analogous to the way agent-computer interfaces helped software engineering agents (as cited via Yang et al., 2024). The paper's catalog of limitations of DOM-based, screenshot-based, and API-based agents is clear and well-documented, and its stakeholder-oriented principles (e.g., safety, developer-friendliness, efficiency) are a useful structuring device. The distinction between AWIs as stateful interfaces and MCP as a stateless protocol is a valuable conceptual contribution. However, the central claim that AWIs are 'essential' and will 'directly fix' current problems is not backed by any implementation, controlled experiment, or formal derivation; the paper itself concedes this in Section 1. The value here is as a proposal and a call to action, not as a demonstrated result.

major comments (3)
  1. [Section 3, first paragraph] The central claim that AWIs are "essential for web agent performance" is the load-bearing assertion of the paper, but it is supported only by a qualitative list of mechanisms in Section 3.1. Every mechanism (tailored action spaces, efficient state representations, task queues) presupposes that the interface is the binding constraint on agent performance; the opposite hypothesis—that grounding, planning, or training on existing interfaces would yield equal or larger gains—is not tested or even formulated. A controlled comparison between the same agent using an AWI-like representation (e.g., a compact structured state plus a restricted action set) and the same agent using a DOM, screenshot, or API interface on a fixed set of tasks would make the claim falsifiable. As written, the word "essential" is an unsupported modal claim. I recommend either softening it to a research hypothesis or committing to a concrete experimental protocol that could in principle falsify it.
  2. [Sections 3.1 and 3.3] The sentence "Developing AWIs will directly fix many of the issues highlighted above" asserts a causal relationship that none of the cited evidence demonstrates. For instance, the claim that tailored browser state representations will "significantly reduce the computational cost" (Section 3.1) is plausible but unmeasured; a DOM tree can exceed 1M tokens, but a compact representation could also lose decision-relevant information, and the net effect on task success is unknown. Similarly, "agentic task queues" and "progressive information transfer" (Section 3.3) are design sketches with no latency, throughput, or cost numbers. Because the paper explicitly provides no prototype (Section 1), the causal claims are not backed by the manuscript's own evidence. I would replace "will directly fix" with "could address" and add a short evaluation agenda with specific metrics (task success rate, token cost, wall-clock time, server load) and baseline conditions.
  3. [Sections 3.2 and 4] The proposal's feasibility depends on website owners and developers implementing, deploying, and maintaining AWIs, but the only incentive offered is that AWIs "would avoid development effort towards managing resources due to higher traffic from agents" (Section 3.1). This incentive is not quantified, and it is in tension with the "Developer-friendly" principle (Section 3.2), which requires that AWIs "work seamlessly with the architecture and infrastructure of the website" without disrupting existing services. The paper does not address the cost of developing a new interface, the risk that a standard will fragment, or the fact that many websites operate on thin margins and may not adopt an interface whose benefits to them are unmeasured. For a position paper, a more explicit treatment of the adoption path—for example, a minimal viable AWI that can be layered onto existing pages, or a non-profit consortium mechanism—would make the argument more credible. Without this, the proposal remains a wish-list rather than a plan.
minor comments (6)
  1. [Section 2.2] There is a grammatical typo: "on a online shopping website" should be "on an online shopping website"; the same phrase appears in Section 3.1.
  2. [Section 3.2, Table 1] The principle is named "Optimal representations" in Section 3.2 but Table 1 uses "Optimally-represented"; the inconsistency could confuse readers.
  3. [Section 4, AI Safety paragraph] "they will used in progressively more critical and sensitive scenarios" should read "they will be used in progressively more critical and sensitive scenarios."
  4. [Section 5, MCP discussion] "PostgresSQL" should be "PostgreSQL."
  5. [References and formatting] Several author names contain spurious spacing artifacts (e.g., "Sta ´nczak"), likely due to LaTeX accent handling; these should be cleaned up in the final version.
  6. [Section 2.1, Representational limitations] The phrase "token-efficient representations" is informal; specifying "token-efficient for current LLMs" would be more precise, since token efficiency is model-dependent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a position proposal with no fitted parameters, no derived predictions, and no load-bearing self-citation chain.

full rationale

This paper is an argumentative position piece, not a derivation. It introduces the concept of an Agentic Web Interface by definition and then argues, on qualitative grounds, that interfaces designed for agents could address representational, resource, safety, and action-space limitations of browser UIs and web APIs. There are no fitted parameters, benchmark numbers produced, or 'predictions' that reduce by construction to inputs. The claim that AWIs 'will directly fix many of the issues' is an unsupported causal assertion, but unsupported assertion is an evidence and correctness concern, not circularity: the paper explicitly disclaims a prototype ('we explicitly do not provide a prototype or specific implementation details', Section 1) and offers no implementation. The one formal element, the sequential decision-making framework, is adopted from external work (Zhou et al.) and does no work in deriving the AWI proposal. Self-citations (WebLinX, AgentRewardBench, SafeArena, BrowserGym) appear as examples or as supporting evidence that existing interface workarounds generalize poorly, but the same claims are also supported by external citations (e.g., Deng et al., WebArena, VisualWebArena, Li et al., Pan et al.) and by the paper's own descriptive analysis of DOM/screenshot/API limitations. No uniqueness theorem is imported, no ansatz is smuggled in via citation, and no known result is renamed as a derivation. The central proposal is normative and would require future empirical validation, but it does not reduce to its own inputs.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The paper introduces no free parameters or physical entities. Its central proposal rests on three unverified domain assumptions about interface limitations, stakeholder adoption, and design feasibility.

assumptions (3)
  • domain assumption Current human-facing web interfaces fundamentally limit web agents and cannot be adequately improved by better agent design.
    This is the motivating premise of the paper, argued in Section 2 and used in Section 3.1 to justify AWIs, but it is not proven.
  • domain assumption Website owners will have sufficient incentive to adopt and maintain AWIs and will participate in a standard-setting process.
    The concrete suggestions in Section 3.3, such as task queues and access control, require deployment by website owners; the paper does not analyze adoption incentives or enforcement.
  • domain assumption A standardized AWI can be designed that serves all stakeholders without degrading the human web experience.
    The six guiding principles in Section 3.2 assume compatibility between agent needs and human needs, but the paper does not test or prototype this compatibility.
invented entities (1)
  • Agentic Web Interface (AWI)
    purpose: A proposed standardized interface for web agents to interact with websites in a stateful, efficient, and safe manner.
    The AWI is a conceptual proposal with no implementation, prototype, or falsifiable prediction; the paper explicitly states it does not provide one.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Build the web for agents, not agents for the web." pith.science (2026). https://pith.science/paper/34MQPQNV

@misc{pith2026250610953,
  author       = {Pith},
  title        = {Pith review of: Build the web for agents, not agents for the web},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/34MQPQNV}},
  note         = {Machine review of arXiv:2506.10953}
}
read the original abstract

Recent advancements in Large Language Models (LLMs) and multimodal counterparts have spurred significant interest in developing web agents -- AI systems capable of autonomously navigating and completing tasks within web environments. While holding tremendous promise for automating complex web interactions, current approaches face substantial challenges due to the fundamental mismatch between human-designed interfaces and LLM capabilities. Current methods struggle with the inherent complexity of web inputs, whether processing massive DOM trees, relying on screenshots augmented with additional information, or bypassing the user interface entirely through API interactions. This position paper advocates for a paradigm shift in web agent research: rather than forcing web agents to adapt to interfaces designed for humans, we should develop a new interaction paradigm specifically optimized for agentic capabilities. To this end, we introduce the concept of an Agentic Web Interface (AWI), an interface specifically designed for agents to navigate a website. We establish six guiding principles for AWI design, emphasizing safety, efficiency, and standardization, to account for the interests of all primary stakeholders. This reframing aims to overcome fundamental limitations of existing interfaces, paving the way for more efficient, reliable, and transparent web agent design, which will be a collaborative effort involving the broader ML community.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Agentic Web: Weaving the Next Web with AI Agents

    cs.AI 2025-07 conditional novelty 3.0 of 10

    A position paper defines the Agentic Web as the next web era and proposes a three-dimensional conceptual framework for understanding and building it.

Reference graph

Works this paper leans on

71 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Andriushchenko, A

    M. Andriushchenko, A. Souly, M. Dziemian, D. Duenas, M. Lin, J. Wang, D. Hendrycks, A. Zou, Z. Kolter, M. Fredrikson, E. Winsor, J. Wynne, Y. Gal, and X. Davies. Agentharm: A benchmark for measuring harmfulness of llm agents, 2025. URL https://arxiv.org/abs/2410.09024

    arXiv 2025

  2. [2]

    Introducing claude, Mar

    Anthropic. Introducing claude, Mar. 2023. URL https://www.anthropic.com/news/introducing-claude

    work page 2023

  3. [3]

    Introducing the model context protocol, Nov

    Anthropic. Introducing the model context protocol, Nov. 2024. URL https://www.anthropic.com/news/model-context-protocol

    work page 2024

  4. [4]

    BehnamGhader, V

    P. BehnamGhader, V. Adlakha, M. Mosbach, D. Bahdanau, N. Chapados, and S. Reddy. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961, 2024

    arXiv 2024

  5. [5]

    Boisvert, M

    L. Boisvert, M. Bansal, C. K. R. Evuru, G. Huang, A. Puri, A. Bose, M. Fazel, Q. Cappart, J. Stanley, A. Lacoste, A. Drouin, and K. Dvijotham. Doomarena: A framework for testing ai agents against evolving security threats, 2025 a . URL https://arxiv.org/abs/2504.14064

    arXiv 2025

  6. [6]

    Boisvert, M

    L. Boisvert, M. Thakkar, M. Gasse, M. Caccia, T. L. S. D. Chezelles, Q. Cappart, N. Chapados, A. Lacoste, and A. Drouin. Workarena++: Towards compositional planning and reasoning-based common knowledge work tasks, 2025 b . URL https://arxiv.org/abs/2407.05291

    arXiv 2025

  7. [7]

    Caccia, M

    M. Caccia, M. Thakkar, L. Boisvert, T. L. S. de Chezelles, A. Pich \'e , N. Chapados, A. Drouin, M. Gasse, and A. Lacoste. Fine-tuning web agents: It works, but it's trickier than you think. In NeurIPS 2024 Workshop on Open-World Agents, 2024. URL https://openreview.net/forum?id=SkwtxEkst2

    work page 2024

  8. [8]

    H. Cai, Y. Li, W. Wang, F. Zhu, X. Shen, W. Li, and T.-S. Chua. Large language models empowered personalized web agents. In Proceedings of the ACM on Web Conference 2025, pages 198--215, 2025

    2025

Show all 71 references
  1. [9]

    T. L. S. D. Chezelles, M. Gasse, A. Drouin, M. Caccia, L. Boisvert, M. Thakkar, T. Marty, R. Assouel, S. O. Shayegan, L. K. Jang, X. H. Lù, O. Yoran, D. Kong, F. F. Xu, S. Reddy, Q. Cappart, G. Neubig, R. Salakhutdinov, N. Chapados, and A. Lacoste. The browsergym ecosystem for...

    2025 arXiv

  2. [10]

    X. Deng, Y. Gu, B. Zheng, S. Chen, S. Stevens, B. Wang, H. Sun, and Y. Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36: 0 28091--28114, 2023

    2023

  3. [11]

    Drouin, M

    A. Drouin, M. Gasse, M. Caccia, I. H. Laradji, M. Del Verme, T. Marty, L. Boisvert, M. Thakkar, Q. Cappart, D. Vazquez, et al. Workarena: How capable are web agents at solving common knowledge work tasks? arXiv preprint arXiv:2403.07718, 2024

    2024 arXiv

  4. [12]

    Furuta, K.-H

    H. Furuta, K.-H. Lee, O. Nachum, Y. Matsuo, A. Faust, S. S. Gu, and I. Gur. Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023

    2023 arXiv

  5. [13]

    X. Ge, C. Xu, D. Misaki, H. R. Markus, and J. L. Tsai. How culture shapes what people want from ai. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI ’24, page 1–15. ACM, May 2024. doi:10.1145/3613904.3642660. URL http://dx.doi.org/10.1145/3613904.3642660

    2024

  6. [14]

    Function calling with the gemini api, May 2025

    Google. Function calling with the gemini api, May 2025. URL https://ai.google.dev/gemini-api/docs/function-calling

    2025

  7. [15]

    B. Gou, R. Wang, B. Zheng, Y. Xie, C. Chang, Y. Shu, H. Sun, and Y. Su. Navigating the digital world as humans do: Universal visual grounding for gui agents, 2025. URL https://arxiv.org/abs/2410.05243

    2025 arXiv

  8. [16]

    Y. Gu, K. Zhang, Y. Ning, B. Zheng, B. Gou, T. Xue, C. Chang, S. Srivastava, Y. Xie, P. Qi, H. Sun, and Y. Su. Is your llm secretly a world model of the internet? model-based planning for web agents, 2025. URL https://arxiv.org/abs/2411.06559

    2025 arXiv

  9. [17]

    I. Gur, H. Furuta, A. Huang, M. Safdari, Y. Matsuo, D. Eck, and A. Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023

    2023 arXiv

  10. [18]

    K. Guu, K. Lee, Z. Tung, P. Pasupat, and M.-W. Chang. Realm: Retrieval-augmented language model pre-training, 2020. URL https://arxiv.org/abs/2002.08909

    2020 arXiv

  11. [19]

    H. He, W. Yao, K. Ma, W. Yu, Y. Dai, H. Zhang, Z. Lan, and D. Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. arXiv preprint arXiv:2401.13919, 2024

    2024 arXiv

  12. [20]

    L. Jang, Y. Li, D. Zhao, C. Ding, J. Lin, P. P. Liang, R. Bonatti, and K. Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks. arXiv preprint arXiv:2410.19100, 2024

    2024 arXiv

  13. [21]

    W. G. JSON-RPC. Json-rpc 2.0 specification, Jan. 2013. URL https://www.jsonrpc.org/specification

    2013

  14. [22]

    Karpukhin, B

    V. Karpukhin, B. Oğuz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. tau Yih. Dense passage retrieval for open-domain question answering, 2020. URL https://arxiv.org/abs/2004.04906

    2020 arXiv

  15. [23]

    Khattab and M

    O. Khattab and M. Zaharia. Colbert: Efficient and effective passage search via contextualized late interaction over bert, 2020. URL https://arxiv.org/abs/2004.12832

    2020 arXiv

  16. [24]

    J. Y. Koh, R. Lo, L. Jang, V. Duvvur, M. C. Lim, P.-Y. Huang, G. Neubig, S. Zhou, R. Salakhutdinov, and D. Fried. VisualWebArena : Evaluating Multimodal Agents on Realistic Visual Web Tasks , June 2024 a . URL http://arxiv.org/abs/2401.13649. arXiv:2401.13649 [cs]

    2024 arXiv

  17. [25]

    J. Y. Koh, S. McAleer, D. Fried, and R. Salakhutdinov. Tree search for language model agents, 2024 b . URL https://arxiv.org/abs/2407.01476

    2024

  18. [26]

    H. Lai, X. Liu, I. L. Iong, S. Yao, Y. Chen, P. Shen, H. Yu, H. Zhang, X. Zhang, Y. Dong, and J. Tang. Autowebglm: A large language model-based web navigating agent, 2024. URL https://arxiv.org/abs/2404.03648

    2024 arXiv

  19. [27]

    Lee, M.-W

    K. Lee, M.-W. Chang, and K. Toutanova. Latent retrieval for weakly supervised open domain question answering, 2019. URL https://arxiv.org/abs/1906.00300

    2019 arXiv

  20. [28]

    I. Levy, B. Wiesel, S. Marreed, A. Oved, A. Yaeli, and S. Shlomov. St-webagentbench: A benchmark for evaluating safety and trustworthiness in web agents, 2024. URL https://arxiv.org/abs/2410.06703

    2024 arXiv

  21. [29]

    u ttler, M. Lewis, W.-t. Yih, T. Rockt \

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K \"u ttler, M. Lewis, W.-t. Yih, T. Rockt \"a schel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33: 0 9459--9474, 2020

    2020

  22. [30]

    W. Li, W. Bishop, A. Li, C. Rawles, F. Campbell-Ajala, D. Tyamagundlu, and O. Riva. On the effects of data scale on ui control agents. In Neural Information Processing Systems, 2024. URL https://api.semanticscholar.org/CorpusID:270285816

    2024

  23. [31]

    Z. Liao, J. Jones, L. Jiang, E. Fosler-Lussier, Y. Su, Z. Lin, and H. Sun. Redteamcua: Realistic adversarial testing of computer-use agents in hybrid web-os environments, 2025. URL https://arxiv.org/abs/2505.21936

    2025

  24. [32]

    E. Z. Liu, K. Guu, P. Pasupat, T. Shi, and P. Liang. Reinforcement learning on web interfaces using workflow-guided exploration. In International Conference on Learning Representations ( ICLR ) , 2018. URL https://arxiv.org/abs/1802.08802

    2018 arXiv

  25. [33]

    J. Liu, Y. Song, B. Y. Lin, W. Lam, G. Neubig, Y. Li, and X. Yue. Visualwebbench: How far have multimodal llms evolved in web page understanding and grounding? arXiv preprint arXiv:2404.05955, 2024 a

    2024 arXiv

  26. [34]

    X. Liu, H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, Y. Gu, H. Ding, K. Men, K. Yang, S. Zhang, X. Deng, A. Zeng, Z. Du, C. Zhang, S. Shen, T. Zhang, Y. Su, H. Sun, M. Huang, Y. Dong, and J. Tang. Agentbench: Evaluating llms as agents, 2023. URL https://arxiv.org/abs/2308.03688

    2023 arXiv

  27. [35]

    X. Liu, T. Zhang, Y. Gu, I. L. Iong, Y. Xu, X. Song, S. Zhang, H. Lai, X. Liu, H. Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024 b

    2024 arXiv

  28. [36]

    X. H. L \`u , Z. Kasner, and S. Reddy. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930, 2024

    2024

  29. [37]

    X. H. Lù, A. Kazemnejad, N. Meade, A. Patel, D. Shin, A. Zambrano, K. Stańczak, P. Shaw, C. J. Pal, and S. Reddy. Agentrewardbench: Evaluating automatic evaluations of web agent trajectories, 2025

    2025

  30. [38]

    Murty, C

    S. Murty, C. Manning, P. Shaw, M. Joshi, and K. Lee. Bagel: Bootstrapping agents by guiding exploration with language, 2024. URL https://arxiv.org/abs/2403.08140

    2024 arXiv

  31. [39]

    Murty, H

    S. Murty, H. Zhu, D. Bahdanau, and C. D. Manning. Nnetnav: Unsupervised learning of browser agents through environment interaction in the wild, 2025. URL https://arxiv.org/abs/2410.02907

    2025 arXiv

  32. [40]

    Nakano, J

    R. Nakano, J. Hilton, S. Balaji, J. Wu, L. Ouyang, C. Kim, C. Hesse, S. Jain, V. Kosaraju, W. Saunders, X. Jiang, K. Cobbe, T. Eloundou, G. Krueger, K. Button, M. Knight, B. Chess, and J. Schulman. Webgpt: Browser-assisted question-answering with human feedback, 2022. URL http...

    2022 arXiv

  33. [41]

    Introducing chatgpt search, Oct

    OpenAI. Introducing chatgpt search, Oct. 2024. URL https://openai.com/index/introducing-chatgpt-search/

    2024

  34. [42]

    Introducing deep research, Feb

    OpenAI. Introducing deep research, Feb. 2025. URL https://openai.com/index/introducing-deep-research/

    2025

  35. [43]

    Y. Pan, D. Kong, S. Zhou, C. Cui, Y. Leng, B. Jiang, H. Liu, Y. Shang, S. Zhou, T. Wu, and Z. Wu. Webcanvas: Benchmarking web agents in online environments, 2024. URL https://arxiv.org/abs/2406.12373

    2024 arXiv

  36. [44]

    Patel, M

    A. Patel, M. Hofmarcher, C. Leoveanu-Condrei, M.-C. Dinu, C. Callison-Burch, and S. Hochreiter. Large language models can self-improve at web agent tasks, 2024. URL https://arxiv.org/abs/2405.20309

    2024 arXiv

  37. [45]

    Prasad, A

    A. Prasad, A. Koller, M. Hartmann, P. Clark, A. Sabharwal, M. Bansal, and T. Khot. Adapt: As-needed decomposition and planning with language models, 2024. URL https://arxiv.org/abs/2311.05772

    2024 arXiv

  38. [46]

    Z. Qi, X. Liu, I. L. Iong, H. Lai, X. Sun, W. Zhao, Y. Yang, X. Yang, J. Sun, S. Yao, T. Zhang, W. Xu, J. Tang, and Y. Dong. Webrl: Training llm web agents via self-evolving online curriculum reinforcement learning, 2025. URL https://arxiv.org/abs/2411.02337

    2025 arXiv

  39. [47]

    Y. Qin, S. Hu, Y. Lin, W. Chen, N. Ding, G. Cui, Z. Zeng, Y. Huang, C. Xiao, C. Han, Y. R. Fung, Y. Su, H. Wang, C. Qian, R. Tian, K. Zhu, S. Liang, X. Shen, B. Xu, Z. Zhang, Y. Ye, B. Li, Z. Tang, J. Yi, Y. Zhu, Z. Dai, L. Yan, X. Cong, Y. Lu, W. Zhao, Y. Huang, J. Yan, X. Ha...

    2023

  40. [48]

    Y. Qin, Y. Ye, J. Fang, H. Wang, S. Liang, S. Tian, J. Zhang, J. Li, Y. Li, S. Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025

    2025 arXiv

  41. [49]

    Rawles, A

    C. Rawles, A. Li, D. Rodriguez, O. Riva, and T. Lillicrap. Androidinthewild: A large-scale dataset for android device control. Advances in Neural Information Processing Systems, 36: 0 59708--59728, 2023

    2023

  42. [50]

    Rawles, S

    C. Rawles, S. Clinckemaillie, Y. Chang, J. Waltz, G. Lau, M. Fair, A. Li, W. Bishop, W. Li, F. Campbell-Ajala, et al. Androidworld: A dynamic benchmarking environment for autonomous agents. arXiv preprint arXiv:2405.14573, 2024

    2024 arXiv

  43. [51]

    Schick, J

    T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom. Toolformer: Language models can teach themselves to use tools, 2023. URL https://arxiv.org/abs/2302.04761

    2023 arXiv

  44. [52]

    P. Shaw, M. Joshi, J. Cohan, J. Berant, P. Pasupat, H. Hu, U. Khandelwal, K. Lee, and K. N. Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36: 0 34354--34370, 2023

    2023

  45. [53]

    T. Shi, A. Karpathy, L. J. Fan, J. Z. Hern \'a ndez, and P. Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, 2017. URL https://api.semanticscholar.org/CorpusID:34953552

    2017

  46. [54]

    Shinn, F

    N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao. Reflexion: Language agents with verbal reinforcement learning, 2023. URL https://arxiv.org/abs/2303.11366

    2023 arXiv

  47. [55]

    R. W. Shirey. Internet Security Glossary, Version 2 . RFC 4949, Aug. 2007. URL https://www.rfc-editor.org/info/rfc4949

    2007

  48. [56]

    Sodhi, S

    P. Sodhi, S. Branavan, Y. Artzi, and R. McDonald. Step: Stacked llm policies for web actions. arXiv preprint arXiv:2310.03720, 2024

    2024 arXiv

  49. [57]

    Y. Song, F. Xu, S. Zhou, and G. Neubig. Beyond browsing: Api-based web agents, 2025. URL https://arxiv.org/abs/2410.16464

    2025 arXiv

  50. [58]

    Tiwary, V

    N. Tiwary, V. Dongre, S. A. Chawla, A. Lamani, and D. Hakkani-Tür. From context to action: Analysis of the impact of state representation and context on the generalization of multi-turn web navigation agents, 2024. URL https://arxiv.org/abs/2410.23555

    2024 arXiv

  51. [59]

    Trabucco, G

    B. Trabucco, G. A. Sigurdsson, R. Piramuthu, and R. Salakhutdinov. Towards internet-scale training for agents. In Will Synthetic Data Finally Solve the Data Access Problem?, 2025. URL https://openreview.net/forum?id=6YFuaxXYxP

    2025

  52. [60]

    A. D. Tur, N. Meade, X. H. Lù, A. Zambrano, A. Patel, E. Durmus, S. Gella, K. Stańczak, and S. Reddy. Safearena: Evaluating the safety of autonomous web agents, 2025. URL https://arxiv.org/abs/2503.04957

    2025 arXiv

  53. [61]

    X. Wan, J. Yang, and J. Xiao. Towards an iterative reinforcement approach for simultaneous document summarization and keyword extraction. In Proceedings of the 45th annual meeting of the association of computational linguistics, pages 552--559, 2007

    2007

  54. [62]

    C. H. Wu, J. Y. Koh, R. Salakhutdinov, D. Fried, and A. Raghunathan. Adversarial attacks on multimodal agents. arXiv e-prints, pages arXiv--2406, 2024

    2024

  55. [63]

    T. Xie, D. Zhang, J. Chen, X. Li, S. Zhao, R. Cao, T. J. Hua, Z. Cheng, D. Shin, F. Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems, 37: 0 52040--52094, 2024

    2024

  56. [64]

    T. Xue, W. Qi, T. Shi, C. H. Song, B. Gou, D. Song, H. Sun, and Y. Su. An illusion of progress? assessing the current state of web agents, 2025. URL https://arxiv.org/abs/2504.01382

    2025

  57. [65]

    R. Yan, X. Wan, J. Otterbacher, L. Kong, X. Li, and Y. Zhang. Evolutionary timeline summarization: a balanced optimization framework via iterative substitution. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval, ...

    2011

  58. [66]

    J. Yang, H. Zhang, F. Li, X. Zou, C. Li, and J. Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023

    2023 arXiv

  59. [67]

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, and O. Press. Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems, 37: 0 50528--50652, 2024

    2024

  60. [68]

    Zhang, X

    H. Zhang, X. Liu, and J. Zhang. Summit: Iterative text summarization via chatgpt. arXiv preprint arXiv:2305.14835, 2023

    2023 arXiv

  61. [69]

    Zheng, B

    B. Zheng, B. Gou, J. Kil, H. Sun, and Y. Su. Gpt-4v (ision) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024

    2024 arXiv

  62. [70]

    Zheng, M

    B. Zheng, M. Y. Fatemi, X. Jin, Z. Z. Wang, A. Gandhi, Y. Song, Y. Gu, J. Srinivasa, G. Liu, G. Neubig, and Y. Su. Skillweaver: Web agents can self-improve by discovering and honing skills. 2025. URL https://api.semanticscholar.org/CorpusID:277634081

    2025

  63. [71]

    S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y. Bisk, D. Fried, U. Alon, and G. Neubig. WebArena : A Realistic Web Environment for Building Autonomous Agents , Apr. 2024. URL http://arxiv.org/abs/2307.13854. arXiv:2307.13854 [cs]

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.