Modeling Distinct Human Interaction in Web Agents

Despite rapid progress in autonomous web agents, human involvement remains essential for shaping preferences and correcting agent behavior as tasks unfold. However, current agentic systems lack a principled understanding of when and why humans intervene, often proceeding autonomously past critical decision points or requesting unnecessary confirmation. In this work, we introduce the task of modeling human intervention to support collaborative web task execution. We collect COWCORPUS, a dataset of 400 real-user web navigation trajectories containing over 4,200 interleaved human and agent actions. We identify four distinct patterns of user interaction with agents – hands-off supervision, hands-on oversight, collaborative task-solving, and full user takeover. Leveraging these insights, we train language models (LMs) to anticipate when users are likely to intervene based on their interaction styles, yielding a 61.4–63.4% improvement in intervention prediction accuracy over base LMs. Finally, we deploy these intervention-aware models in live web navigation agents and evaluate them in a user study, finding a 36.8% increase in user-rated agent usefulness. Together, our results show structured modeling of human intervention leads to more adaptive, collaborative agents.

Introduction. Recent advances in large language models (LLMs) have enabled AI agents to perform increasingly complex tasks in web navigation (Deng et al., 2024; Shi et al., 2017; Yao et al., 2022; Zhou et al., 2023). Despite this progress, effective use of such agents continues to rely on human involvement to correct misinterpretations or realign behavior with user preferences (Amershi et al., 2014; Misra et al., 2017; Saunders et al., 2017). However, current agentic systems lack an understanding of when and why humans intervene. As a result, agents may pursue autonomy under incorrect assumptions about user intent, overlooking critical factors as tasks unfold (Hadfield-Menell et al., 2016; Mitchell et al., 2025). Even when agents proactively stop to check with users, they often do so at inappropriate moments or interrupt too frequently with unnecessary confirmation requests (Chen et al., 2024; Huq et al., 2025), forcing users to step in mid-execution and incurring a heavy oversight burden (Bansal et al., 2024; Wang et al., 2020, 2025). Modeling when humans are likely to intervene can help agents anticipate preventable mistakes, minimize unnecessary disruptions, and reduce the oversight burden without sacrificing reliability.

To build more effective collaborative agents, autonomy should complement human involvement rather than override it, and engage users only when their input is necessary. Although recent work has explored proactive assistance in collaborative agent (Feng et al., 2024; Huq et al., 2025; Ramrakhya et al., 2025; Shao et al., 2025), these approaches typically focus on specific interaction mechanisms, including asking follow-up questions or co-planning with users, rather than modeling the broader spectrum of human interaction patterns that arise during execution, such as mid-task intervention, alternative action-taking, and transfer of control.

Related work. Autonomous Web Agents Web automation has been transformed by LLM-based agents capable of navigating complex environments. Benchmarks such as Mind2Web (Deng et al., 2024) and WebArena (Zhou et al., 2023) have pushed agents toward real-world, multi-domain tasks that use HTML and accessibility tree representations. The emergence of recent Computer Use capabilities from models like Claude (Anthropic, 2024) and Operator have further closed the gap between human browsing and machine execution as demonstrated by plugin-based web agents tools like WebCanvas (Pan et al., 2024), WebOlympus (Zheng et al., 2024), OpenWebAgent (Iong et al., 2024), and Taxy (TaxyAI, 2024) that can integrate into natural browsing contexts.

However, these extensions often prioritize autonomy over collaboration, lacking mechanisms for interactive control by users. Our work builds upon this plugin-based paradigm and emphasizes human-agent collaborations beyond solo agent autonomy.

Modeling Human-Agent Collaboration Human-AI collaboration has been widely studied in various settings, ranging from robotics (Ajoudani et al., 2018; Chandrasekaran and Conrad, 2015) and productivity tools (Khadpe et al., 2020; Zhang et al., 2021) to LLM-based collaboration. Frameworks such as Magentic-UI (Mozannar et al., 2025), Cocoa (Feng et al., 2024), Collaborative Gym (CoGym) (Shao et al., 2025), and Collaborative STORM (Shao et al., 2024), A2C (Tariq et al., 2024) further advance these interactions by introducing mechanisms for co-planning and co-execution across real-time web and persistent document environments. By moving toward non-turn-taking protocols, these systems enable more flexible control where humans and agents can operate in the same execution space. Notably, LLM-based collaboration has made significant progress in writing assistance, with examples including CoAuthor (Lee et al., 2022), PEER (Schick et al., 2022), VISAR (Zhang et al., 2023). Early interactive systems like PUMICE (Li et al., 2019) and PLOW (Allen et al., 2007) demonstrated the value of end-user programming and demonstration. Previous studies such as interaction to impact (Zhang et al., 2025), TrustAgent (Hua et al., 2024), and ToolEmu (Ruan et al., 2024) focus primarily on safety and trustworthiness. We shift our focus towards the overall communication patterns in human-agent web browsing collaboration.

Method. This motivates a central question: Can agents proactively anticipate when human intervention is likely and adapt their behavior accordingly? To answer this, we introduce COW- CORPUS: a real-user corpus of 400 humanagent collaborative task execution trajectories on the open web. The dataset comprises over 2,748 agent action steps and 1,476 human action steps, with step-level annotations marking when users intervened by pausing, resuming, or overriding agent execution. Analyzing COWCORPUS reveals that human involvement is driven by three recurring needs: error correction, preference refinement, and assistive takeover. Although these motivations appear in individual actions, they combine into consistent higher-level collaboration strategies over the course of a task. Accordingly, we group users into four distinct collaboration styles: Takeover, Hands-on, Hands-off, and Collaborative; which capture how users balance supervision, intervention, and share control with the agent (§3).

Building on this empirical characterization, we formulate the task of modeling human intervention patterns conditioned on user collaboration style. We cast this problem as a stepwise sequence prediction task: at each agent action, the model estimates the likelihood of user intervention given the evolving task context. We train LMs in two settings: (i) a general intervention-aware model that moves beyond purely autonomous execution, and (ii) style-conditioned models adapted to specific collaboration styles. Across multiple model backbones, our interventionaware models improve intervention prediction accuracy by 61.4–63.4% over baselines (§4).

Beyond offline prediction accuracy, we integrate our intervention-aware models into live web agents and evaluate them through a user study on real-world web tasks. Agents equipped with intervention modeling achieve a 36.8% increase in user-perceived usefulness over baseline systems, demonstrating that anticipating human intervention leads to more adaptive and effective human–agent collaboration in practice (§5).

We formulate the human-agent collaboration as a Partially Observable Markov Decision Process (POMDP). Given a task instruction q, the two actors, z∈{agent, human} following policies πagent and πhuman, attempt to complete it by taking a sequence of discrete actions τ= [a1, a2, . . . , aT], where each action atat time tis taken either by the agent or the human. At each time step t, we construct a multimodal observation ot= (Vt, At) of the current state, consisting of the current screenshot Vtand the webpage accessibility tree At. This observation is passed into the actor policy to generate the next action at= π(ot|τ0:t−1), where τt−1 = [(o1, a1), . . . , (ot−1, at−1)] denotes the past trajectory. By default, the agent generates the proposed action ˆat= πagent(o0:t, a0:t−1).

At any time step, the human can choose to intervene, formalized as a binary intervention variable yt∈{0, 1}. We define this human intervention modeling task to be a step-wise binary classification, where the objective is to learn a predictive model fθthat estimates:

We approach this using a large multimodal model (LMM) optimized via supervised finetuning (SFT). The model takes a serialized prompt containing (1) the history trajectory τt−1, (2) the current observation ot, and (3) a description of the agent-proposed action ˆat. The model is fine-tuned to generate designated tokens: <ask_user> or <agent_continue>, indicating the intervention decision to intervene or proceed.

We introduce COWCORPUS, a dataset for studying human intervention patterns in collaborative web workflows. In this section, we first describe the data collection process (§3.1), then analyze the motivation behind user intervention (§3.2) and summarize user interaction patterns (§3.3).

To ensure COWCORPUS is consistent with established benchmarks and reflects individual user preferences, we designate a mixture of free-form tasks and benchmark tasks in our dataset — (1) 10 standard tasks from the Mind2Web dataset (Deng et al., 2024) and (2) 10 free-form tasks of the participants’ own choice.

Discussion. More broadly, this work suggests a shift from optimizing agent autonomy to designing agents that dynamically adapt to human preferences and collaboration styles over time.

Conclusion. In this work, we show that human intervention in web navigation constitutes a structured behavioral signal that reflects distinct collaboration styles, ranging from passive supervision to active co-piloting. We introduce COWCORPUS, a dataset of 400 real-user web navigation trajectories designed to support the study of intervention modeling in collaborative settings. Our analysis reveals that while proprietary generalist models demonstrate strong reasoning capabilities, they struggle to capture the temporal dynamics of when users choose to intervene. By fine-tuning models on collaborative interaction traces, we bridge this gap, achieving a 61.4– 63.4% improvement in intervention prediction accuracy over base LMs. Finally, we deploy our intervention-aware models in a live web agent, and show that anticipating human intervention leads to tangible benefits in practice, increasing user satisfaction by 36.8%. Together, these results highlight the value of deliberately modeling human-agent interaction patterns. We hope that our work encourage the development of agents that are more responsive, adaptive, and capable of functioning as truly collaborative partners.