A service can give an agent a dedicated tool and still be reachable through the ordinary website. That leaves product owners with two routes to understand when deciding what an agent may do.
Grace Huckins’s November 2 MIT Technology Review report examines that choice. One approach lets agents navigate pages through browser interactions. Another gives them connectors that expose selected service functions. The reporting also considers who controls discovery and transactions as agents become intermediaries between people and online businesses.
The technical distinction predates this report. In a June 2025 position paper, Xing Han Lù and colleagues proposed interfaces designed specifically for agents. Their suggestions include restricted access to sensitive information, confirmation before destructive actions, and ways for a person to pause and redirect a task. The authors explicitly offer a design argument rather than a prototype or a completed standard.
Their discussion of internal APIs identifies a failure to watch for: elevated credentials can bypass safeguards in the human interface. A connector’s existence alone tells a buyer little about its authority.
Can this agent draft a change, submit it or commit it? An enterprise service needs a clear answer across the routes it supports. If a person stops a task, the team should be able to establish what happened before the stop and whether any work remains queued.
The new reporting describes a changing market, not proof that one interface makes an agent safe. A practical evaluation should follow a representative task through its available routes and inspect the resulting state. The product decision is which actions the service exposes, who authorizes them and how the user takes control again.