Tools

Human-readable files that map UI elements to stable names, making simulations robust against front-end changes.

Two-way translator between agent and browser, parsing HTML and translating high-level actions into browser commands.

Robust browser automation using Playwright for interacting with live web interfaces.

Interprets current UI state into stable, structured representation for agent consumption.

Perplexity.AI

ArXiv, https://arxiv.org.

Wikipedia is a free online encyclopedia, created and edited by volunteers around the world.

Classic web search interface with SERP layout, search suggestions, and comprehensive result listings.

E-commerce platform interface for product search, filtering, reviews, and purchase simulation.

Conversational AI interface with message-based interaction, context preservation, and multi-turn dialogue.

Creates new search queries based on task descriptions or words found in previously seen documents.

Decides which search results to click on based on their rank and snippet text.

Decides when to end the search session based on simple rules, like number of relevant documents found.

Manages search sessions, tracking interactions and maintaining session state throughout the simulation.

Advanced query generation using large language models for natural language understanding and contextual query expansion.

Dynamic stopping criteria that adapts based on search progress, user satisfaction, and information gain metrics.

Interprets the UI state using Recipes to understand the current interface configuration.

Creates high-level strategies for achieving search goals and managing information needs.

Performs specific actions like clicking, typing, or navigating based on current plans.

Reviews past actions and improves strategy based on outcomes and learning.

Generates new questions, sub-goals, and exploratory directions during search.

Specialized agent for academic research tasks with advanced literature analysis and synthesis capabilities.

E-commerce focused agent that helps with product discovery, comparison, and purchase decision support.

Straightforward policy operating on hand-crafted IF-THEN rules and numerical thresholds.

Uses LightGBM classifier to decide which action component to use based on learned user behavior patterns.

Advanced LLM-based policy that reasons about situations and provides structured responses with explicit thoughts.

Evaluates current simulation state including user persona, information needs, and interaction history.

Specialized orchestration policy for academic research workflows with emphasis on systematic exploration and validation.

Shopping-focused orchestration policy that balances exploration, comparison, and decision-making for purchase tasks.

News and content consumption policy that balances breadth, depth, and credibility in information gathering.

Natural language description-based personas that capture user behavior patterns, preferences, and search strategies through conversational AI.

Structured JSON-based personas with demographic data, job information, background, and detailed behavioral parameters for precise simulation control.

Algorithmic personas with quantified behavioral rules like frustration levels, patience thresholds, and decision-making parameters for predictable simulation outcomes.

Advanced persona management system that combines LLM, JSON, and rule-based approaches for comprehensive user behavior simulation with adaptive intelligence.