Anthropic and Google reveal different control models for AI agents

Structural differences in reach

Claude Duet functions as a standalone client available on the web, iOS, and Android. It operates by running tasks on a virtual machine with its own browser, which allows it to continue working even if your laptop is turned off.

In contrast, Gemini Intelligence is designed as an integrated layer within the Android operating system. Rather than being an app you must open, it lives inside the surface you are already using, such as providing context for a grocery list or a travel brochure.

As OrcaRouter reported, the two products offer different types of reach. Anthropic highlights a breadth of over 6,000 connectors and plugins, while Google defines its reach as a boundary where the assistant can only access apps the user explicitly allows.

Who initiates the action

The two agents differ significantly in how they start tasks. Claude Duet is designed to be proactive, sending morning briefs and drafting next steps, but it throttles these actions to manage plan limits.

Gemini Intelligence is described by Google as primarily reactive for automation tasks. It only acts on direct user commands and stops immediately once the task is finished, providing live notifications of its progress.

Google's framework emphasizes explicit user control as a core principle. While it includes proactive features like Magic Cue, the primary automation surface is built to wait for a trigger rather than running in an idle loop.

The gate before an action

Anthropic provides a specific list of actions that will trigger a pause in Claude Duet. Before the agent sends a message, makes a payment, signs a document, or deletes anything, it must show the exact action and wait for user approval.

Gemini Intelligence uses a design-based gate rather than an enumerated list of verbs. It requires user confirmation before making purchases and relies on granular opt-in controls for different features and app-level automations.

The practical difference lies in where the agent stops. Duet stops between a draft and its consequence for four specific types of actions. Gemini stops at the end of a task or before a purchase, with permissions deciding if the task runs at all.

Refusals and measurements

Anthropic is transparent about what it refuses outright, such as trading securities or moving money between accounts. It also refuses credit card numbers and government IDs at the message layer, meaning these never even reach the approval prompt.

Google does not publish a list of actions that are refused even with user consent. This creates a difference in auditability, as a user can read the specific boundaries of Anthropic's product but cannot see the full scope of Google's refusals.

Neither company provides independent benchmarks to prove how well these agents actually perform tasks. Anthropic shares usage limits based on plan costs, while Google focuses on its rollout schedule and structural permission models.

Availability and rollout

The release plans for these agents are currently asymmetrical. Claude Duet is a research preview for Pro and Max subscribers in the US, with specific exclusions for Team and Enterprise plans.

Google is executing a staged rollout across the Android platform. It began with the latest Samsung Galaxy and Google Pixel phones this summer, with plans to expand to watches, cars, and glasses later this year.

Because neither company provides a device-by-device inventory for today, users must look at the rollout waves. One vendor provides a clear list of what it won't do, while the other provides a clear map of where it is going.

Source origins and data limits

The information regarding Gemini Intelligence is derived from two specific Google posts released on May 12, 2026. These include a primary announcement by Mindy Brooks, the VP of Product Management, and a corresponding security post by Dave Kleidermacher, the VP of Platforms Security and Privacy.

Neither of these posts establishes which specific devices carry which features as of October 24, 2026. Consequently, this comparison does not claim that any particular phone currently possesses these capabilities, as the shipped inventory is not listed in the official documentation.

Infrastructure and external context

For those building their own agent loops, OrcaRouter provides an alternative to fixed vendor apps by serving more than 200 models behind a single OpenAI-compatible endpoint and key. This service passes through the provider's list price with 0% markup and includes automatic failover between different models.

While the agents discussed here are consumer surfaces with pre-set gates, developers can use models like Claude Opus 5.5 as a line item. This allows for custom approval logic where the model executes an approved action rather than being tied to a specific vendor's application boundaries.