WebAgents proposal separates agent creation from hosting and execution
A research paper describes a signed package format that lets agent workflows move between runtimes while users keep control of models, permissions, and budgets.
In plain words
AI agents are programs that plan tasks and use tools on a user's behalf. The WebAgents paper proposes packaging an agent's behavior in a signed file so it can run on different systems without the user handing over control of their accounts or resources. In tests, the system blocked injected rule violations while running unchanged workflows on different model setups.
What was announced
Researchers introduce WebAgents, a system that unbundles how agents are created, hosted, and executed. The core is a signed agent.pkg file carrying portable workflows, skills, schemas, and resource requirements, so agent behavior can move across creators, hosts, and runtimes without tying publication to one execution environment or transferring authority over user resources. Hosts distribute the public program and may keep private services server-side.
At run time, an OpenClaw-based browser and runtime verifies the package, obtains authorization, and binds it to user-controlled models, tools, permissions, and budgets. The authors demonstrate three unrelated workflows executed unchanged under local and API-backed models, including a literature review with live services and a website-published travel agent. In a controlled suite, the system blocked all 716 injected contract violations while accepting all 295 matched benign controls, with median pre-execution processing below 0.5 seconds per package. The paper outlines a path toward NANDA-based discovery and interoperable package ecosystems.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, threat model, and evaluation details are not available for verification.
- The security results come from the authors' own controlled test suite and have not been independently confirmed.
For security and IT teams
- Scope
- The findings describe a research prototype evaluated in a controlled suite with three workflows, not a deployed production system.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.