On October 21, Anthropic opened Claude Duet’s waitlist, starting its research preview with US Pro and Max subscribers. It announced separate iOS and Android apps and web access. According to the company, the assistant waits for approval before sending messages, paying, signing documents or deleting data.
As of October 22, 2026, this is an analysis of official materials, not our hands-on review. The practical question is whether an assistant reduces the number of things you must remember without creating a new queue of checks. The answer will depend on real work rather than the persuasiveness of a product presentation.
Access and resource use: what to check first
The FAQ excludes Team and Enterprise for now. Background checks consume Claude’s shared allowance, while larger tasks require more model work. Duet inherits the account’s memory, skills and connected apps. The company also warns that it may misunderstand conversations or judge urgency incorrectly.
Our assessment: do not buy a subscription solely expecting immediate access to the new assistant. Readers in Ukraine should check eligibility, not just recognise a plan name. During a trial, record how much ordinary Claude work remains available after delegating tasks. This makes it easier to judge whether the benefit justifies the resources used.
What makes an agent different from a chat
Duet uses its own cloud computer and browser, even with your device off. Connected accounts are checked periodically, not continuously.
Announced use cases include morning briefs, finding unpaid invoices, drafting reminders, and preparing documents, presentations or dashboards for user review.
Anthropic describes agents through a cycle of planning, acting, observing and adjusting. Behaviour depends on the model, operating rules, tools and environment. The company treats access to a calendar and permission to send an invitation as different user decisions.
That provides a useful evaluation framework. A well-written plan does not prove execution. A task needs an observable result: a prepared document, a checked collection of data or a reply draft. When work involves another service, look for confirmation in that service. Otherwise, intention, preparation and completion can easily become confused.
Security involves several separate mechanisms
The Managed Agents engineering article explains separating credentials from the code-execution environment, with external tools accessible through a proxy and secure vault. This describes Anthropic’s architectural approach, not an independent audit of every Duet operation.
Anthropic’s Claude Code auto mode article describes screening unwanted instructions and classifying actions before execution. It acknowledges a trade-off between false blocks and missed risks. The research results should not automatically be treated as Duet reliability measurements.
Its browser-agent research explains prompt injection: third-party pages or documents can contain commands intended to redirect an assistant’s task. Anthropic explicitly says the problem remains unsolved. Having protective measures does not guarantee error-free operation.
Our practical recommendation is to begin with limited access and tasks whose mistakes are easy to correct. A useful first example would be gathering meeting material and preparing questions without sending anything to participants.
Also examine approval requests: are the recipient, content and consequences clear? A generic yes without that context does little to preserve meaningful control.
Connecting a service also means moving data
Anthropic’s current privacy policy explains that integrations can receive inputs, outputs and instructions to perform actions. Third-party services apply their own data practices. Some connections remain active until disabled; the policy advises checking your authority to grant access.
For work email or client documents, first agree within the team which data is appropriate to use. A small test set is a better starting point than an entire archive. Define who reviews the result, where the final version belongs and what happens when a conclusion is incomplete or contradictory. A new application cannot replace that process design.
How to assess value without marketing expectations
We suggest testing three tasks of different complexity: preparing a short brief from supplied material, comparing two documents and drafting a fact-based response. Write down a completion criterion for each in advance. Afterwards, measure not only time saved but also time spent correcting, clarifying and checking sources.
Assess the usefulness of the assistant’s messages separately. A discrepancy caught in time matters more than a long report with no next step. If every result requires reconstructing the entire history yourself, delegation has not worked yet. If you can quickly verify the result and make a decision, try the same workflow with different material.
Our conclusion is to evaluate Claude Duet as an experiment in delegation, not a ready-made employee replacement. Repeatable results in your own tasks will be the strongest evidence in its favour. Control over actions beyond text preparation remains the key boundary. Until independent testing is available, the manufacturer’s promises should be treated as promises.
Comments