Anthropic released Claude Duet as a research preview on October 21, 2026. It is a proactive AI agent — built on Opus 5.5 — that monitors connected applications, delivers a morning brief summarizing pending tasks and relevant information, and initiates actions like scheduling and invoice handling without waiting to be prompted. The design choice that distinguishes it from a reactive assistant: Duet asks for human approval before taking high-stakes actions, particularly financial or legal ones. It is currently waitlisted for US Claude Pro and Max subscribers and available in the Claude mobile app and on claude.ai. Team and Enterprise tiers are not included in the initial rollout.
Duet enters a field where competitors are moving fast. Meta’s Muse and OpenAI’s Dots are positioned in similar territory. The differentiator Anthropic is emphasizing is the governance architecture: proactive but gated, not proactive and autonomous.
stateDiagram-v2 direction TB state "Idle: monitoring connected apps" as Idle state "Task or trigger identified" as Task state "Executes directly" as Direct state "Requests human approval" as Request state "Approved — executes" as Approved state "Declined — cancels" as Declined state "Logs action and returns" as Log [*] --> Idle Idle --> Task : trigger detected Task --> Direct : low-stakes action Task --> Request : high-stakes action Request --> Approved : human approves Request --> Declined : human declines Direct --> Log Approved --> Log Declined --> Log Log --> Idle
For engineers: the authorization boundary is the hard architectural problem
Proactive agents differ from reactive assistants in a way that matters structurally. A reactive assistant waits for a prompt, operates within a session, and its errors are local to that interaction. A proactive agent initiates, persists across sessions, and its errors can compound across time before anyone notices.
The engineering challenge is defining and enforcing the authorization boundary: which actions the agent can take without human input, which require explicit approval, and which are never delegated regardless of user preference. Anthropic’s design for Duet answers this for one product in one consumer context. Enterprise deployments face a harder version of the same problem: the authorization boundary is organization-specific, varies by role, and needs to be enforceable across multiple agents from multiple vendors.
Approval fatigue is the technical failure mode to design against. If the agent requests approval for low-stakes actions frequently enough, users begin approving without reading. That makes the approval gate a compliance artifact rather than a real control. The systems that will hold up in practice are the ones that correctly classify the stakes of an action before deciding whether to act or ask — and that classification logic is where the engineering complexity actually lives.
The audit trail question is non-negotiable for enterprise deployment. Every action an agent takes, approved or autonomous, needs to be logged in a format that supports review. Not because incidents are common, but because when they happen, the timeline of what the agent did and when is the first thing legal and compliance will request.
For business owners: “the agent asked me” is not a governance model
Claude Duet’s approval-request pattern is well-designed for a consumer product. It is incomplete as an organizational governance framework.
If an employee approves an agent action that turns out to be wrong — an invoice sent to the wrong party, a meeting booked that reveals a confidential schedule, a communication sent before a decision was finalized — the approval does not transfer responsibility away from the organization. The agent asked and the human approved. Both the agent’s design and the human’s decision sit inside your accountability boundary as the employer and the entity that chose to deploy the tool.
This is not a reason to avoid proactive agents. It is a reason to define your authorization framework before your employees start using them on work tasks. Some already are: if anyone on your team subscribes to Claude Pro or Max, they have access to Duet’s waitlist. You may not have decided anything about governance, but they are deciding with their approval clicks.
The conversations that need to happen are not complicated: which tasks in our workflows are well-specified enough to be handled by a proactive agent, what decisions require human sign-off and why, and who owns the review if something goes wrong. These are not new governance questions. They are the same questions you would ask about any new automated system in your business. Proactive AI agents make them urgent because the deployment is already in progress.
My take: the organizations that define boundaries before deployment stay ahead of incidents
At a class-action settlement administration company I worked with for several years, one of the highest-impact projects I ran was an automated returns processing and fulfillment system, built with tight integration to USPS APIs to handle the routing and execution of complex mail-return workflows. The system processed mail at high volume with close to 100 percent accuracy — and it did so because we designed the decision tree before we wrote a single line of code.
That upfront work was architecture-level thinking: what does the system do when it encounters an ambiguous return? What are the escalation triggers? What gets logged and where? What requires a human decision before the workflow continues? We answered those questions before implementation, not after a production incident forced the conversation. The project succeeded because the boundaries were defined, documented, and tested in advance.
The same logic applies to proactive AI agents in enterprise settings. Anthropic has made reasonable design choices in Duet’s approval architecture for a consumer product. The enterprise version of that governance model is your responsibility to build before the tool arrives — not after it has been running in your employees’ workflows for three months. The organizations that get ahead of this will have cleaner incident histories than those that discover the boundaries when something goes wrong.