WebRider: A New AI Agent for Live-Web Assistance with Intent Contracts
A recent paper on arXiv (2608.06704) presents WebRider, a hierarchical AI framework aimed at enhancing live-web assistance through the formalization of delegated policies as 'intent contracts.' This innovative study identifies a significant shortcoming in existing live-web agents, which are assessed solely on their final outputs, neglecting policy adherence. A comprehensive live audit demonstrated that while a robust controller successfully completes 99.2% of tasks, it complies with policy constraints in only 38.8% of instances, indicating that task completion does not guarantee fidelity. WebRider counters this issue by keeping a detailed operational record of goals, constraints, evidence obligations, answer formats, and task-specific persona controls. Its architecture features a top-layer controller for contract maintenance, a middle layer for executing guarded actions, and likely a lower layer for execution. This research, authored by a team of researchers, contributes to AI and web automation by proposing a novel method to ensure compliance with user-defined policies during web tasks.
Key facts
- WebRider is introduced in arXiv paper 2608.06704.
- Current live-web agents are evaluated only on final answers, ignoring policy constraints.
- A strong controller completes 99.2% of tasks but honors all policy constraints in only 38.8% of cases.
- WebRider formalizes delegated policy as an 'intent contract'.
- Intent contract includes goals, constraints, evidence obligations, answer form, and task-local persona controls.
- WebRider uses a hierarchical architecture with a top-layer controller and a middle layer for guarded executable actions.
- The paper is announced as a new type on arXiv.
- The audit reveals a critical gap in live-web assistance.
- WebRider aims to bridge the gap between task completion and policy fidelity.
- The paper is available at https://arxiv.org/abs/2608.06704.
Entities
Institutions
- arXiv