PrivAgentFlow: Agentic Workflow for Distributed Privacy Control in Web Agents

Autonomous web agents increasingly operate on sensitive personal and contextual information, yet most privacy-preserving frameworks rely on static access policies or centralized filters that fail to adapt to task dynamics, execution context, or user intent. We introduce PrivAgentFlow, an agentic workflow framework that formulates privacy preservation as a distributed, governable optimization process embedded within the agent’s decision flow. Each node in the workflow enforces the data minimization principle by jointly deciding what information to expose and where execution should occur (local vs. API), balancing privacy risk, task relevance, and computational cost. This composition of locally adaptive nodes yields a workflow that is self-regulating, transparent, and dynamically aligned with the assigned privacy policies. In large-scale web-agent evaluations, PrivAgentFlow reduces environment-based privacy leakage by 15.5\%, API-exposure leackage by 92.5\%, and improves utility by 2.3\% across 84 web tasks, establishing a scalable foundation for trustworthy and distributed privacy governance in web-native autonomous agents.