WebOperator

WebOperator: Action-Aware Tree Search for Autonomous Agents in Web Environment

🌐 Website • 📃 Paper • 💻 Code • 🗂️ Trajectories • 🧪 WebReplayBench

WebOperator

📖 Abstract

LLM-based agents often operate in a greedy, step-by-step manner, selecting actions solely based on the current observation without considering long-term consequences or alternative paths. This lack of foresight is particularly problematic in web environments, which are only partially observable—limited to browser-visible content (e.g., DOM and UI elements)—where a single misstep often requires complex and brittle navigation to undo. Without an explicit backtracking mechanism, agents struggle to correct errors or systematically explore alternative paths. Tree-search methods provide a principled framework for such structured exploration, but applying them to the web introduces two challenges: replaying an action sequence may not reconstruct the same page, and some actions may irreversibly change the website. To address these challenges, we introduce WebOperator, a tree-search framework that enables validated backtracking and strategic exploration. Our method incorporates a best-first search strategy that ranks actions by both reward and reversibility, along with a robust backtracking mechanism that verifies reconstructed pages before continuing from them. It prioritizes reversible actions, postpones potentially irreversible ones, and discards earlier branches when its post-execution heuristic flags a possible persistent change. Experiments on WebArena and WebVoyager demonstrate the effectiveness of the framework. On WebArena, it achieves 54.6% task success with gpt-4o, outperforming all prior tree-search methods using the same backbone. Controlled ablations show that naive tree search can reduce performance, whereas action-aware selection and validated backtracking improve corrective search. We further introduce WebReplayBench, a web environment that records which actions actually change the website’s state, to evaluate whether WebOperator detects irreversible actions and avoids repeating them during backtracking.

📊 Results on WebArena Benchmark

Agent Model Overall (#812) Reddit (#106) GitLab (#180) Shopping (#187) CMS (#182) Map (#109) Multisite (#48)
BrowserGym gpt-4 15.0 20.2 19.0 17.2 14.8 25.5 -
LM-TS gpt-4o 19.2 11.3 13.9 27.8 16.5 26.6 16.7
Go-Browse qwen-2.5-7b 22.6 30.7 15.3 22.4 25.3 17.9 -
AWM gpt-4 35.5 50.9 31.8 30.8 29.1 43.3 -
Branch-n-Browse gpt-4o 35.8 50.9 36.7 34.6 26.4 46.8 18.8
WebPilot gpt-4o 37.2 65.1 39.4 36.9 24.7 33.9 -
AgentOccam gpt-4-turbo 45.7 67.0 43.3 46.2 38.9 52.3 16.7
AgentSymbiotic claude-3.5 52.1 66.0 51.0 48.0 49.0 60.0 29.0
ScribeAgent gpt-4o 53.0 73.7 59.7 45.8 37.9 56.3 -
                 
WebOperator gpt-4o 54.56 76.42 52.78 49.20 54.95 55.05 31.25

Experimental trajectories: run logs, search trees and step-by-step viewers for every task.

📂 Project Structure

.
├── weboperator/                 # Source code for the web agent
│   └── configs/                 # Agent configurations (demo, WebArena, WebVoyager)
├── webshepherd/                 # Source code for the Process Reward Model
├── browsergym/                  # Source code for the web environment simulator
├── gobrowse/                    # Source code for the experience retrieval module
├── webarena/, webvoyager/       # Benchmark task files and evaluation code
├── webreplaybench_eval/         # Harness that runs WebOperator on WebReplayBench
├── utils/                       # Scripts that collect runs and compute benchmark scores
├── run.py, demo.py              # Entry points
└── README.md

# The project page lives on the gh-pages branch.

⚙️ Installation

1️⃣ Clone the repository

git clone https://github.com/kagnlp/WebOperator.git
cd WebOperator

2️⃣ Create environment

conda create -n weboperator_env python=3.12
conda activate weboperator_env
# or using pip and virtualenv (if python3.12 is the system python)
python -m venv weboperator_env
source weboperator_env/bin/activate  # On Windows use `weboperator_env\Scripts\activate`

3️⃣ Install dependencies

Refer to the Running with Docker section if you don’t have admin rights to install Playwright dependencies.

pip install -r requirements.txt
playwright install chromium --with-deps # Need admin rights

4️⃣ Set up environment variables

Create a .env file by copying the example configuration:

cp .env.example .env

Then open the .env file and update any necessary values (such as API keys, website urls) according to your environment.

🚀 Usage

Run the Demo

python demo.py

or

python run.py --config weboperator/configs/demo.yml

🐳 Running with Docker

Useful if you don’t have admin rights to install Playwright dependencies. No need to create a virtual environment or install dependencies.

docker compose run --user $(id -u) weboperator --config weboperator/configs/demo.yml

Skeleton Code

Boilerplate code (demo.py) to run WebOperator on an interactive, open-ended task:

import gymnasium as gym
import browsergym.core  # register the openended task as a gym environment
from weboperator.tree_search_agent import TreeSearchAgent
from weboperator.action_generator import ActionGenerator
from weboperator.models.openrouter import OpenRouterModel

# start an openended environment
env = gym.make(
    "browsergym/openended",
    task_kwargs={"start_url": "https://map.google.com/"},  # starting URL
    wait_for_user_message=True,  # wait for a user message after each agent message sent to the chat
    headless=False
)

# Create an agent
action_generator = ActionGenerator(
    model=OpenRouterModel("openai/gpt-oss-20b:free")  # Set OPENROUTER_API_KEYS in .env file
)
agent = TreeSearchAgent(
        chat_mode=True,
        action_generator= action_generator,
    )

# run the environment <> agent loop until termination
obs, info = env.reset()
while True:
    preprocessed_obs = agent.obs_preprocessor(obs) # Preprocess observation
    action = agent.get_action(preprocessed_obs, env) # Decide action
    obs, reward, terminated, truncated, info = env.step(action) # Act and Observe
    if terminated or truncated:
        break
# release the environment
env.close()

Sample Output

Open-ended + Google Maps

Screenshot

🎯 Benchmarks

WebArena Setup

Before running WebArena experiments, you must host the WebArena websites and configure the corresponding endpoints.

Host Websites (choose one):

Set Environment Variables:

PUBLIC_HOSTNAME=<YOUR_SERVER_DOMAIN_OR_IP>

export WA_SHOPPING=http://${PUBLIC_HOSTNAME}:7770
export WA_SHOPPING_ADMIN=http://${PUBLIC_HOSTNAME}:7780/admin
export WA_REDDIT=http://${PUBLIC_HOSTNAME}:9999
export WA_GITLAB=http://${PUBLIC_HOSTNAME}:8023
export WA_GITLAB_IP=${PUBLIC_HOSTNAME}
export WA_WIKIPEDIA=http://${PUBLIC_HOSTNAME}:8888/wikipedia_en_all_maxi_2022-05/A/User:The_other_Kiwix_guy/Landing
export WA_MAP=http://${PUBLIC_HOSTNAME}:3000

Inference

Run the agent on each benchmark using the corresponding configuration file.

Evaluation

Move the inference outputs and compute benchmark scores.

WebReplayBench

WebReplayBench is a separate benchmark: a website whose database records every persistent change, with backtracking scenarios as a JSON dataset. The harness in webreplaybench_eval/ runs WebOperator on it, to measure how often backtracking replays destructive actions, from naive replay up to full WebOperator. With WebReplayBench cloned next to this repository, on a Docker host:

bash webreplaybench_eval/run_ladder.sh 8

⚙️ Agent Configuration Explanation

Environment

env:
  task_type: "openended" # ["webarena", "webvoyager", "openended"]
  max_steps: 100 # Maximum steps per episode (For BrowserGym)
  headless: false # false: show browser UI; true: hide browser UI

Experiment

experiment:
  results_dir: "./results/openended/gpt-oss-20b" # Directory to save results. Give relative path.

Agent

agent:
  allow_unauthorized_page: true # Whether allow visit to pages outside the benchmark domain

Models

models: # List of models used in the agent
  action_model: # Unique identifier of the model
    type: "OpenRouterModel" # Options: ["OpenAIModel", "AzureOpenAIModel", "OpenRouterModel", "OpenHFModel"]
    model_name: "openai/gpt-oss-20b:free"
  reward_model:
    type: "AzureOpenAIModel"
    model_name: "gpt-4o"
    temperature: 1.0

Agent Components

components:
  action_validator: # Optional: Action validator configuration
    allow_invalid_action: false # Whether to allow semantically invalid actions (Default: false)
    allow_invalid_page: false # Whether to allow navigation to invalid pages (Default: false)

  observation_processor: # Observation processor configuration
    optimized: true # true: use full or visible-only observation based on the observation size. false: always use visible-only observation
    truncate_error_message: true # Truncate long error messages

  action_processor: # Action processor configuration
    merge_strategy: "sum" # ["sum", "max", "none"]: strategy to merge semantically similar actions. "none": do not merge.
  
  recovery_assistant: # Optional: Recovery assistant configuration
    recover_from_invalid_page: true # true: forcefully go_back or tab_close when on invalid page
    recover_from_captcha: true # Whether to allow human intervention for captcha recovery

  backtrack_manager: # Optional: Enables backtracking mechanism
    destruction_aware: true # Whether to re-root the tree after executing destructive actions
    simulation_verified: true # Whether to do snapshot-validation or not

  action_selector: # Action selection strategy configuration
    selection_strategy: "action-aware" # options: ["highest-reward", "action-aware"]
    search_budget: 4 # Frontier budget
    n_candidates: 2 # Number of solution candidates to consider
    max_depth: 20 # Maximum search depth
    max_steps: 20 # Maximum steps (excluding backtracking steps)

  rephraser: # Optional: Enables instruction rephraser
    model: "action_model" # Model used for rephrasing instructions

  retriever: # Optional: Enables examples retriever (Set RETRIEVER_API_SERVER in environment variables)
    type: "faiss" # ["faiss", "bm25"]
    model: "all-MiniLM-L6-v2" # Sentence transformer model (for faiss retriever)
    top_k: 5 # Number of examples to retrieve

  judge: # Reward and checklist model configuration. Note: Applicable only for multiple action candidates
    prompt_type: "web_operator"  # Options: likert_scale, web_shepherd, web_operator
    checklist_model: "reward_model" # Model used for checklist generation
    reward_model: "reward_model" # Model used for reward estimation

  action_generator:
    max_retry: 5 # Maximum retries for generating syntactically and semantically valid actions
    full_action_space: # List of all possible actions
      - "click"
      - "fill"
      - "select_option"
      - "goto"
      - "go_back"
      - "go_forward"
      - "scroll"
      - "new_tab"
      - "tab_focus"
      - "tab_close"
      - "stop"
    action_space_type: "adaptive" # options: ["fixed", "adaptive"]
    candidates: # List of action generator candidates
      - name: "simple_action_generator" # Unique name for the candidate
        model: "action_model" # Model to use 
        history_length: 5 # Number of previous steps to include in the context
        rephraser: false # Whether to include rephrased task instruction
        retriever: false # Whether to include retrieved examples
      - name: "action_generator_w_retriever"
        model: "action_model"
        history_length: 3
        rephraser: false
        retriever: true
      - name: "action_generator_w_rephraser"
        model: "action_model" 
        history_length: 4
        rephraser: true
        retriever: false

📝 Citation

Please cite our paper:

@article{dihan2025weboperator,
  title={WebOperator: Action-Aware Tree Search for Autonomous Agents in Web Environment},
  author={Dihan, Mahir Labib and Hashem, Tanzima and Ali, Mohammed Eunus and Parvez, Md Rizwan},
  journal={arXiv preprint arXiv:2512.12692},
  year={2025}
}
This site is open source. Improve this page.