Skip to content

WebAgents: splitting agent creation, hosting, and execution for the agent-native Web

WebAgents: Unbundling Agent Creation, Hosting, and Execution for an Agent-Native Web

arXiv: 其他 AI 安全论文· arXiv:2611.21978· Dario Fenoglio, Martin Gjoreski, Marc Langheinrich, Pradyumna Chari, Ramesh Raskar·Indexed by this site · 2 days ago· Originally publishedDate not givenDefenses· Its place in the scan record →
AI summary

Researchers propose WebAgents, which separates the creation, hosting, and execution of agents into independent roles, targeting an Agentic Web composed of independently developed agents and services. The system uses signed agent.pkg to package portable workflows, skills, schemas, and resource requirements; an agent-native browser/runtime based on OpenClaw verifies the package and obtains authorization, binding at runtime to user-controlled models, tools, permissions, and budgets; the host distributes public programs and can retain private services on the server side. Three mutually unrelated workflows run unmodified under local and API models, including a literature review using live services and a travel agent published to a website. In a controlled test suite, WebAgents blocked all 716 injected contract violations while allowing all 295 matching benign controls, and the median time for automatic processing before execution of the three packages was below 0.5 seconds each. The authors also outline a path toward a NANDA-based discovery and interoperable package ecosystem, while retaining execution authority in the user runtime.

Read originalarxiv.org