Connector or browser action: which should an AI agent use?

Why connectors come first for account data

A connector calls a service operation through the backend. It returns named fields and provider status values, avoids layout changes, and can apply a permission rule to each operation.

• Fewer page loads and fewer moving parts.
• Clear read and write boundaries.
• Provider responses can be used as completion evidence.

When the browser is the right tool

The browser is appropriate when the request concerns the open page, when a service has no suitable API operation, or when visual layout itself carries the information. Full-page screenshots can help with rendered content that cannot be extracted cleanly as text.

A practical routing order

Explicit instructions win. Page-specific work stays in the page. Requests about your inbox, workspace or analytics account use a connected service when available. Ambiguous requests should be clarified instead of guessed.

Different tools fail in different ways

Browser actions can fail when a layout changes, a modal covers the target or the active tab no longer matches the plan. Connector calls can fail because a scope is missing, an account lacks access or a provider rejects the request.

• Keep the provider's real error instead of replacing it with a generic failure.
• Do not silently switch from a denied connector write to browser clicking.
• After a write, read back the resulting state whenever the service supports it.

Examples of the routing decision

“What emails arrived today?” clearly refers to personal account data and should use Gmail when it is connected. “Copy the code block from this page” is tied to the active page and should use the browser. “Find five Budapest software companies” is open-web research and should start in the background rather than opening five tabs.

Sources and further reading