SIAgentNews.com
← All news
DevelopingAgentsResearch

WebAgents proposes splitting who builds, hosts and runs an AI agent

A new arXiv paper separates agent creation, hosting and execution, so a signed package can run under the user's own models, permissions and budgets.

· First reported by arxiv.org

716injected contract violations blocked in the controlled test suite

The short version

  • WebAgents splits agent work into three roles: creators, hosts and runtimes (the software that runs the agent).
  • A signed file called agent.pkg carries workflows, skills, schemas and resource requirements.
  • The user's runtime verifies the package, gets authorization, and attaches the user's own models, tools, permissions and budgets.
  • The test suite blocked all 716 injected violations and accepted all 295 matched benign controls.
  • Median automated pre-execution processing stayed below 0.5 s for all three packages.

Why it matters

If agents become common on the web, someone must decide who controls what they can touch and spend. This design keeps that authority with the user instead of the company that publishes the agent. It is an early research proposal, not a product you can use today.

The idea: three jobs, not one

Today, an AI agent (software that plans and takes actions using an AI model) usually comes bundled. The same company often builds it, hosts it and runs it. The WebAgents paper, posted on arxiv.org, argues that these roles should be pulled apart.

The authors have in mind an "Agentic Web": a web of agents and services built independently by different people. For that to work, an agent's behavior has to travel between creators, hosts and runtimes. Publishing an agent should not tie it to one place where it runs, and it should not hand over control of a user's resources.

How the package works

The core piece is a signed file called agent.pkg. It carries the agent's portable workflows, its skills, its schemas (descriptions of the data it expects) and a list of the resources it needs.

A browser or runtime built on OpenClaw checks the package and asks for authorization. Only then does it connect the agent to the user's chosen models, tools, permissions and budgets. Hosts hand out the public program, and they can keep private services on their own servers.

What the tests showed

The authors ran three unrelated workflows without changing them, using both local models and models reached through an API. Examples included a literature review using live services and a travel agent published on a website that used the user's selected personal resources.

In a controlled suite, the system blocked all 716 injected contract violations and accepted all 295 matched benign controls. Median automated processing before execution was under 0.5 s for all three packages. These are the authors' own results, not an outside evaluation.

What comes next

The paper ends with a path toward discovery based on NANDA and toward ecosystems of packages that work with each other. The stated goal is to keep execution authority in the user's runtime.

Nothing here is ready for the public yet. It shows one answer to a question that will grow: who is in charge when an agent from someone else runs on your behalf?

What to watch

  • Whether other agent platforms adopt a signed, portable package format like agent.pkg.
  • How NANDA-based discovery would let people find agent packages.

Sources and links

First reported by arxiv.org. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news